Skip to main content

Command Palette

Search for a command to run...

Release Notes

Release notes

Every Cursor product in one timeline, newest first, with up to 10 recent releases each. Filter by product, open a product's page for its full history, or follow along with the RSS feed.

Showing 98 releases from every product

Cursor CLI

Agent

  • Admin Prompt hooks from the team dashboard now run as prompts. Prompt hooks a team admin sets up in the dashboard now run in the CLI with their prompt and model, as they do in the desktop app. Before, the CLI ran each one as an empty command, so the prompt never ran. Dashboard command hooks with no script are now skipped.
  • Sandboxed commands can use /tmp on Linux. Set "privateTmp": true in sandbox.json, with readBoundary set to "workspace", and /tmp inside sandboxed commands becomes the workspace's own scratch folder, the same one $TMPDIR points to. Tools that write to a literal /tmp path stop failing, files stay there for the next command, and other workspaces can't see them. macOS ignores the setting.
  • Skip login scripts in the agent's shell. Set CURSOR_LOGIN_SHELL=0 and the agent starts its shell without running login scripts such as /etc/profile, ~/.bash_profile, or ~/.zprofile, so they can't reset your PATH. Your rc file, such as ~/.bashrc or ~/.zshrc, still runs. Leave it unset to keep the current behavior.

Security

  • Other programs running as you can no longer read the CLI's memory on Linux. The CLI now blocks other processes under your user account, including commands the agent runs, from reading its memory, environment, or open files, or attaching a debugger to it. This keeps the credentials the CLI holds out of their reach. To attach a debugger or tracer, set CURSOR_AGENT_ALLOW_PROCESS_INSPECTION=1 before starting agent.
  • MCP resource downloads follow your file-write rules. When the agent saves an MCP resource to disk, the write goes through the same permission checks and approvals as any other file edit, and it must land inside your workspace. Paths that start with ~, use a URL scheme, or contain .. are refused, and Ask and Plan modes no longer save resources.

Self-hosted workers

  • Self-hosted workers can use MCP servers running locally on your machine. agent worker now runs URL servers from ~/.cursor/mcp.json whose address is localhost, 127.0.0.1, or [::1], such as a desktop app's local MCP endpoint, so agents on the worker can call their tools. Cursor's backend can't reach those addresses. Other URL servers still run on the backend, as described in MCP servers.
  • agent worker --help explains when self-hosted workers exit. A new Lifecycle section says a worker exits with code 0 whenever Cursor releases it: its agent was archived, killed, or expired, its claim went stale, or it sat idle past --idle-release-timeout. It recommends running each pool worker under a supervisor, such as a systemd unit, a Windows scheduled task, or a Kubernetes Deployment, and restarting it with agent worker --pool <name> start --wait. See session lifecycle. Examples in agent worker controller --help now use the command you ran, such as cursor agent.

Fixes

  • Fixes to plan mode canvas writes. In plan mode, the agent can no longer write files outside the canvases folder through a symlinked folder, a hard-linked canvas, or a file name with backslashes on macOS and Linux. Plan mode now refuses canvas writes when a folder inside ~/.cursor/projects is a symlink to another location. Agent mode is unchanged.
  • Fixes to edit previews. Control characters in a file, such as escape codes, now show as visible symbols in the edit preview instead of clearing the screen, renaming the terminal window, or ringing the bell. The file on disk and the text sent to the model are unchanged.
  • Fixes to PowerShell errors on Windows. When a PowerShell command the agent runs fails to parse, such as one using && in Windows PowerShell 5.1, the error now comes back in UTF-8, so localized messages are no longer garbled. Error locations now point at the line in the agent's command instead of a temporary script. This applies to CLI and ACP sessions.
  • Fixes to saving your login in the macOS keychain. The CLI no longer exposes your access token or API key in the process list while it saves your login. A token or API key that is too long or contains a line break now gets a clear error and leaves your saved login unchanged. Two CLI sessions saving at the same time no longer fail with a duplicate item error.
  • Fixes to clicks on computer-use workers. On a worker started with --computer-use, PNG and JPEG screenshots keep the size the agent works in instead of shrinking, so clicks land where the agent aimed instead of up and to the left of the target.

Cursor SDK

  • Admin Local runs say which team MCP allowlist applies to Cursor-hosted MCP servers. When a local agent connects to a Cursor-hosted MCP server, the run prints a [cursor-sdk] line to stderr naming the MCP allowlist in effect, why it applies, and which servers it allows. Each server's line prints once per process, and MCP servers you host yourself never trigger it. Works in TypeScript and Python.
  • Python SDK callers now see SDK notices. The Python SDK prints the SDK's [cursor-sdk] notices, such as which MCP allowlist applies, to your process's stderr, each distinct message once, as the TypeScript SDK does. Other diagnostic output goes to the cursor_sdk.bridge logger at debug level, which you can see with CURSOR_SDK_LOG=debug (see troubleshooting). Python agents also no longer stall when the SDK's background process writes a lot of diagnostic output.
  • Fixes to nested subagents with inherited tools. TypeScript local agents that set local.subagentInherit no longer launch subagents endlessly when no inherited tool can finish the task. A subagent can still start one more of its own, but that child has to answer with the tools it has instead of delegating again.

Cursor CLI

Agent

  • Fewer lost edits when the agent rewrites a file. If you change a file after the agent read all of it or wrote it earlier in the session, the agent's next write that replaces the whole file is refused, nothing is written, and the agent is told to read the file again and retry. Reading only part of a file turns this check off for that file until the agent reads all of it or writes it again.
  • Sandboxed commands run in the right folder. Sandboxed shell commands now start in the directory the agent asked for instead of the workspace root. If the sandbox can't enter that directory, the command fails with a message naming a folder to use, and the agent no longer asks to rerun it outside the sandbox.

Self-hosted workers

  • Add MCP servers to self-hosted workers at launch. Programs that start agent worker can pass extra stdio MCP servers in the CURSOR_WORKER_MCP_CONFIG_JSON environment variable, and the worker loads them when it starts, alongside the servers in ~/.cursor/mcp.json. A disabledServers list in the same JSON turns off any servers you name. The worker removes the variable from its environment, so the commands it runs don't inherit any secrets in it.

Cursor SDK

  • Steer cloud runs from TypeScript and Python. run.steer(text) now delivers text into a running cloud run instead of always resolving revert_to_followup, and Python adds Run.steer and AsyncRun.steer. If the cloud run has already ended, the text is queued as a follow-up and the outcome may include the run it starts, so don't send it again. Accounts without steering access get a FeatureUnavailableError.
  • Token usage for each model call in local runs. In TypeScript, a local run's onDelta callback now receives a model-call-completed interaction update as each model call finishes, with that call's input, output, cache read, cache write, and reasoning tokens. Track cost as it grows during a turn, while turn-ended still carries the turn's total. The ModelCallCompletedUpdate type and ModelCallCompletedUpdateSchema are exported.

Cursor mobile

Voice calls

  • Voice calls keep going past two hours and through server drops. A voice call no longer ends at the two-hour mark. The app moves the call to a fresh connection during a pause, with no reconnect chime, and the conversation carries on. Calls the server drops now reconnect to the same conversation instead of hanging up.
  • Hear Project worker results during voice calls. If a team admin has turned on Projects, a voice call in a Project's chat reads each worker's result aloud once, as soon as that worker finishes. Results keep coming even if you cover or leave the chat during the call.
  • Voice calls can send requests while a plan waits for Build. Asking a question or giving a follow-up during a voice call no longer fails with "The agent needs you in the chat" when the agent's plan is waiting for you to build it. The request goes through like a typed message. Calls still hold your request while the agent waits on an open question or a sign-in prompt.
  • Voice call turns read like typed ones. Messages a voice call sends to the agent are now hidden in every call mode, and the "Sent by voice call" label is gone. Work from a voice turn shows the same Worked sections, replies, and Changes cards as a typed request, and an instruction given mid-turn continues that turn instead of starting a new one. Messages stay hidden after you reopen the chat.
  • Choose which status chimes play. During a voice call, the Settings menu has a Status Chimes submenu with a switch for each chime: Task Sent, Result Arrives, Needs Decision, and Error. Your choices carry over to later calls, starting from your previous chime setting.
  • See which agent has a live voice call. While you're away from its chat, an agent with a voice call in progress shows a green waveform chip with the call timer in the agent list. Tap the row to return. Double-tapping Stop or the mic no longer starts a voice call by accident, and links in a call's transcript now open.

Agents

  • Open Agent Store files and Project notes from chat. Tapping a file path in an Agent Store, such as a Project's Context, or a Project notes link in an agent's reply or tool details opens the file in the in-app markdown preview instead of doing nothing. Links inside a previewed file open their own preview, so you can follow a plan or task note without leaving the chat.
  • Running agent chats open without a blank screen. Opening a chat whose agent is still working shows the recent conversation right away, then fills in the rest of the running turn. Before, a running turn with many steps could keep the chat blank until every step loaded. Turns you expand while it fills in stay expanded.

Fixes

  • Voice calls. If you stay quiet when a call opens while the agent is mid-task, the call now greets you after a short pause instead of staying silent. After a dropped call reconnects, the voice keeps its context without repeating earlier updates. In any chat, an edit you reject no longer shows up in the turn's Changes card.
  • Returning to the app. Coming back to Cursor from the background refreshes your sign-in, team memberships, and plan again, instead of leaving them out of date. The machine and pool list in the composer also updates when you return.

Agents Window

Projects

  • Refresh your repository list in the project picker. The project picker's Repos section has a Refresh button that reloads your cloud repositories, and reopening the picker keeps the current list on screen while it updates in the background. If the list fails to load, the picker says "Couldn't load repositories" instead of "No repos", and it no longer lists the same folder twice.
  • Run a multi-repo selection on your computer. When you pick several repos in the project picker, your computer's row in the Run on menu now offers a recent local workspace that has those repos checked out. If a workspace has only one of them, the row runs the agent there with that repo only, and if none match, the row explains that multi-repo selections run in Cloud only.

Voice

  • Turn voice chat off in Settings. The Voice settings tab has a new Voice chat switch. Turning it off hides voice chat and voice call cards in the Agents Window and ends any active call. The switch is locked with a note when your team admin has turned voice chat off.
  • Choose which voice call chimes play. Voice chat settings now have a Status Chimes section that lists each voice call chime: Task Sent, Result Arrives, Needs Decision, and Error. You can preview each one and turn it on or off separately, and a call that's already running picks up the change.
  • Voice chat explains when your network blocks it. When your network or region blocks the voice service, the Start voice chat button in the agent input is disabled, and its tooltip says why, such as asking IT to allow the voice provider, instead of the call failing to connect. Cursor rechecks on launch, when your network changes, and when the window regains focus.

Chat and composer

  • Start a canvas or side chat from the side rail or an empty panel. The side rail and the empty editor panel now offer Canvas and Side Chat entries, so you can open a new canvas or side chat without going through the + menu. They appear whenever the + menu offers them for the current agent.
  • Watch a browser subagent's browser. When a subagent is using the browser, its Task card now has a Show button that opens that browser in the browser panel. Your cursor stays in the chat input, so you can keep typing.
  • Choose which old chats to delete. The Developer: Delete Old Chats command now opens a dialog where you pick an age (over 30, 60, 90, or 180 days since last update) and whether to delete only archived chats or all chats. The delete button shows how many chats will be removed, and deleted chats stay gone in other open windows.

Files, changes, and pull requests

  • Videos play inline and images load in the PR tab. Videos uploaded to a pull request now play inline with player controls in the PR tab's description and comments, and uploaded images load instead of showing as broken. The activity timeline also uses colored icons and web-matching wording like "approved this pull request", and opening the stack dropdown centers the current PR in long stacks.

Browser and design mode

  • Open Annotate notes survive a page reload. In the browser's Annotate mode (Design Mode), a note you have open now stays open through a page reload and comes back with its unsaved text when the page shows its element again.

Cloud agents

  • Move to Cloud with Include Changes opens pull requests against the right branch. When you move a local agent to the cloud with Include Changes, its pull request now targets your working branch whenever that branch is on the remote, instead of the repo default with an inflated diff. This also works when the moved repo is a secondary repo in a multi-repo environment.
  • Admin MCP servers your admin blocks show as Disabled by admin for cloud agents. When you run an agent in the cloud, the prompt's + menu now shows an MCP server that your team's MCP server settings block as Disabled by admin, instead of a switch.

Customization and settings

  • The Agent SDK skill works in local agents. The Agent SDK create-agent skill now works in local agents as well as cloud agents, and installed Agent SDK skills no longer appear twice.

Fixes

  • Fixes to Projects. The tracked-task board now loads after you join a Project, tracked tasks in a new Project migrate reliably, and the member facepile shows the Project's own members. The empty Projects section no longer flashes at startup, and local chats with side chats are no longer offered for cloud Projects, so those side chats aren't lost.
  • Fixes to errors and New Chat. The branch picker now explains when an IP allowlist blocks the request, and sending to a workspace on an offline self-hosted worker says the worker is offline instead of asking for a repo URL. New Chat defaults to the repo and environment of the agent you last started from New Chat, not whichever agent you were viewing.

Cursor IDE

Agent and chat

  • Edit dictated text without stopping dictation. While you dictate into the agent input, typing no longer cancels dictation. Your edits stay and new words are added after them. Clearing the input still stops dictation.
  • Skill chips show where a skill comes from. Hovering a skill chip in the agent chat input shows its source and repository plus its description trimmed to four lines.
  • Copy or save Mermaid diagrams as images. Mermaid diagrams in chat now have a More actions menu with Copy as Image and Save as Image, so you can paste a diagram into a doc or save it as a PNG.
  • Fixed a misleading Other Models usage entry on weekly plans. On plans with a single weekly usage allowance, the usage status bar and ring in the agent panel no longer show Other Models at 0% when you select a third-party model. They now show a single Cursor Models percentage, whichever model you select.

MCP

  • Agent finds MCP tools by what they do. When the agent searches your MCP tools, it now matches words in each tool's description, not only server and tool names. Searches also report MCP servers that are still starting or failed to connect instead of returning nothing.
  • MCP Apps call their own tools without an approval prompt. Interactive MCP Apps in the agent chat now call tools on their own server directly, without asking you to approve the server or tool first. Tools their server hides from apps stay off-limits.
  • Admin Team MCP allowlist URL rules match servers more precisely. URL rules in your team's MCP allowlist now match a server by its scheme, host, port, and path, so wildcard rules allow or block exactly the servers admins intend in the editor. Deny rules still block everything they blocked before, and a host with a trailing dot, such as example.com., matches the same server as one without.

Customization and settings

  • Admin Private Inference supports rules, remote MCP servers, and clearer gateway errors. With Private Inference, your team rules and User Rules now apply to agent turns, and you can view and edit User Rules in Settings. Remote MCP servers can now connect. Cursor-hosted MCP servers stay off. When your inference gateway or proxy fails, the agent error names it and the cause, and Auto-review follows the reviewer set in your Private Inference config.
  • Open your Origin settings from Source Control settings. Cursor Settings > Source Control has a new Manage Origin Settings row. Its Open button takes you to your Origin and pull request settings on cursor.com.
  • Admin Marketplace plugin installs show progress and admin refusals. Adding a plugin from the marketplace in Customize now shows "Adding..." until the install finishes. If your team admin doesn't allow the plugin, Cursor shows a warning with the reason and marks the plugin Disabled, instead of briefly listing it as added.

Fixes

  • Fixes to MCP. MCP servers that only support the newer stateless handshake now connect, connections cut off mid-response are retried, and stdio servers that print non-JSON text no longer show as errored in MCP settings.
  • Fixes to updates. Update checks and downloads no longer hang when the server stops sending data, and restarting to install an update is no longer disrupted when a newer version appears mid-quit. On Windows, finished updates no longer trigger update recovery at startup, and leftover download folders are cleaned up.
  • Fixes to local agent reliability. Cancelling an agent request now ends it immediately, waking from sleep no longer triggers false extension host hang warnings, and a failed checkpoint restore or redo now shows an error with the reason.
  • Fixes to Plan mode and file protection. In Plan mode, the agent can no longer delete non-markdown files or edit them through file links or symlinks with a markdown name. File-Deletion Protection now asks before deletes in Run Everything mode.
  • Fixes to terminals. Terminal text no longer renders in the wrong font and size with GPU acceleration on. A terminal whose process has exited, including a reattached one, no longer shows as busy or asks for confirmation when you close it.
  • Fixes to agent chat. The Agents > Text Size setting scales chat text again, Cmd+A (Ctrl+A) selects transcript text, and images an agent links with a workspace-relative path or file: URI now load. Rename Chat applies after another window renamed the chat, and Bugbot's Fix in Cursor works when the PR's repo is a git submodule in your workspace.
  • Fixes to remote reconnect. When a remote server restarts and can't resume the previous session, the "Unable to reconnect to the remote workspace" notification now offers Reload Window instead of a Reconnect Now button that could not succeed.

Cursor CLI

Agent

  • Agent questions now include the context behind them. When the agent stops to ask you a question, it now writes what it found and why in its message first. You no longer get a question that refers to reasoning hidden in collapsed thinking.
  • Fixed file-scoped rules and skills going missing after some reads. When the agent reads a file that matches a glob-scoped rule or a nested skill, it now gets that rule or skill even if the file is empty, an image, or too large to read. Before, reading a file like an empty __init__.py dropped the rule silently, and later reads in the same folder didn't bring it back.

Self-hosted workers

  • agent worker shows whether your team allows minting source-control tokens. When you start a self-hosted worker with --mint-scm-token, the status panel and startup summary now say whether the team setting is on, off, unavailable, or couldn't be checked. If it's off, the worker tells you a team admin can turn it on under Cloud Agents → Self-hosted Machines, and any-repo pool workers started without --clone-git-repos get a note that they need it.

Fixes

  • Fixed MCP tool calls with empty arguments. When the agent calls an MCP tool with empty arguments, the call now runs instead of failing with a parse error.
  • Fixes to agent edits that replace every match. When the agent replaces every occurrence of a string in a file, such as stripping indentation or collapsing blank lines, it no longer deletes code outside the matched text while still reporting success.

Cursor CLI

Fixes

  • To-do and question updates no longer fail over a missing closing bracket. When the agent sends its to-do items or its questions for you as a list that stops right after the last complete item, the CLI now closes the list and runs the tool call instead of failing it.

Cursor CLI

Cloud Agents

  • Fixed Cloud Agents starting without your unpushed commits. When you hand a conversation to a Cloud Agent with & from a branch that has local commits but no uncommitted changes, the CLI now pushes those commits so the agent starts from the code you have. Before, the agent started from the remote copy of the branch without telling you. If the push fails, the handoff now stops with an error.

MCP

  • The agent can find MCP tools by what they do. When you have many MCP tools, the agent's tool search now matches words in tool descriptions as well as names, ignoring case, so it can find a tool without knowing its exact name. Search results also show which MCP servers aren't ready yet.

Self-hosted workers

  • Admin --mint-scm-token replaces --mint-github-token for pool workers. agent worker --pool now takes --mint-scm-token to receive short-lived source control tokens during claimed runs, and the help, status panel, and errors now say SCM token. A team admin still has to turn on SCM token minting. The old --mint-github-token flag still works as a hidden alias, so existing worker scripts keep running. See the worker options.
  • Admin Workers using --connect-workspace now download the live workspace server from downloads.cursor.com. agent worker --connect-workspace now fetches the live workspace server from https://downloads.cursor.com first and falls back to https://cursor.blob.core.windows.net if that fails, checking each download's signature as before. On locked-down networks, allow outbound HTTPS to downloads.cursor.com alongside the existing host. agent worker --help names both. See worker networking.
  • Fixes to worker hints. agent worker help text and the desktop sharing and computer use warnings now name the debug command the same way the CLI's other hints do, usually agent worker debug, instead of always saying cursor-agent worker debug.

Cursor SDK

  • Compact a local agent's conversation from Python with agent.summarize(). Agent.summarize() and AsyncAgent.summarize() free up context on demand, like /summarize in the CLI and agent.summarize() in TypeScript. The returned run streams SDKSummaryMessage events, and wait() returns the summary text. Calling it on a cloud agent raises ConfigurationError.
  • Fixes to installs and local runs. The SDK now finds its platform-specific package when installed with pnpm or other layouts that place it next to the SDK. Local runs keep their max mode setting and their full result text, even when it contains NUL characters, and pick up checkpoints that another process writes to the same local store.

Web

  • Secret scanning on push only flags secrets you add. When you push to an Origin repository, secrets that already existed in the repository, for example in a moved or copied file, no longer block or warn on your push. When a push is rejected for test data, the error tells you to skip with -o secret_scanning.skip=used_in_tests rather than splitting or encoding the value.
  • Rename repositories without breaking links or clones. Repository admins can now rename a repository from Settings > Advanced. Old links, clones, and API calls that use the previous name keep redirecting to the repository, and no other repository can take the old name. Repositories that GitHub is the source of truth for show the rename option disabled until you detach them from GitHub.
  • Admin SSH keys move to My settings, and team SSH settings move to Security. Manage the SSH keys you clone and push with from My settings, which you can also reach from the command palette. Team SSH certificate authorities, the certificate requirement, and the IP allowlist now live in the Security section of the Origin settings page instead of an SSH tab under API Keys, and Origin settings show a Your SSH keys row with your key count and a link to manage them.
  • Admin Keep clones small by hiding refs from them. Repository admins can now list ref patterns under Hidden from clones in the Advanced tab of repository settings, and matching branches and tags are left out of normal clones. The default branch is never hidden. A clone prints which refs are hidden, and anyone who needs one can still fetch it by name.
  • Draft pull requests open as drafts again. The Never open PRs as drafts setting has been retired, so a pull request you or your agents ask to open as a draft now opens as a draft, whether from the web, the Origin API, the Origin CLI, or a cloud agent. Moving an open pull request back to draft is no longer refused, and new pull requests still open ready for review unless a draft is requested.
  • Admin Merge rulesets with no branch patterns can be saved and are marked as not applied. In a repository's Settings > Rules and Protections, you can now save a merge ruleset with an empty custom branch list, matching what the Origin API and Origin CLI already accept. The editor warns that the ruleset targets no branches, and the list shows it as No branches with a Not applied badge.
  • Admin Name your Agentic Owners policies. Agentic Owners policies can now set an optional name in policy.yml. On the pull request page, a named rule shows its name as the title, and its owners move into the hovercard, so two rules with the same owners are easy to tell apart. Policies without a name look the same as before.
  • Clones and fetches of mirrored repositories no longer fail on commits Origin hasn't synced yet. In repositories mirrored from GitHub, HTTPS and SSH clones and fetches are now served by Origin by default. When you fetch a commit or pull request ref the mirror hasn't picked up yet, Origin fetches it from GitHub on demand instead of returning an error.

API

  • Admin List repositories across all your namespaces in one Origin API call. Listing hosted repositories can now return every repository you can read across your personal namespace, your teams, shared namespaces, and repositories you collaborate on, with per-namespace counts. You can narrow it to specific namespaces. Access APIs also return each policy's ID and accept it instead of a policy name when you set repository or namespace access. See the Origin API changelog for details.

Fixes

  • Fixes to review inbox loading. The review inbox on the Codebase home shows the inbox right away when you return from a pull request. Pull requests in Origin repositories that mirror out to GitHub no longer drop out of the inbox when their GitHub mirror details are missing, sections with slow filters no longer fail after 30 seconds, and merge, auto-merge, and draft actions keep working after you close an agent's review pane. Inbox search now picks up stacked pull request changes within seconds and no longer shows pull requests from deleted repositories.
  • Fixes to pull request pages and mergeability. Closing a pull request now turns off its auto-merge, team review requests stay pending until a team member other than the author approves or requests changes, and discarding a pending review while it is being submitted no longer deletes it. Mergeability now updates when a namespace access policy changes, no longer fails intermittently when many checks update at once, and no longer shows a head branch as deleted based on another branch. The Activity tab no longer jumps while loading, Code Review Gate checks count toward CI status, late redeliveries no longer send duplicate Slack or email notifications, and failed ownership exemption or break-glass requests explain what went wrong.
  • Fixes to stacked pull requests. On a pull request page, a pending conflict check or restack in the stack now refreshes on its own without a reload. Merging, landing, and reverting in stacks with a long merged history no longer fail on the stack size limit. Adding a pull request to a stack that already has 200 unmerged pull requests is now refused with a clear message instead of breaking the stack.
  • Fixes to git clone, fetch, and push. Clones and fetches of busy repositories retry on another server instead of failing, and a credential without access now gets a 403 permission denied error instead of a 503. The first push to an imported repository no longer changes its default branch, and idle repositories and long download URLs no longer break clone kits.
  • Fixes to GitHub mirroring. Mirrors no longer get stale refs or miss updates when an Origin server restarts, and syncs retry when GitHub briefly rejects fresh credentials.
  • Admin Fixes to rulesets, apps, and access. Removing or uninstalling an app now drops its ruleset bypass right away, saving a ruleset that names an uninstalled app fails with a clear error, and concurrent ruleset edits no longer occasionally fail to apply. The bypass and group pickers in repository settings stay usable while reloading, and the accept-invite page explains when an invite went to a different email. People without access to a repository's checks no longer see check output or details links, and opening a private app you can't read now says it couldn't be found.
  • Admin Fixes to Origin settings. On the Origin settings page, an organization admin who isn't on the team now sees that only team members can see its namespace instead of a load error, and the PR review provider setting no longer greys out while a change saves. Pages that read your team's Origin namespace stop retrying requests that will never succeed.
  • Fixes to Origin API errors. REST API POST requests are no longer cancelled before they run. When an Origin backend service fails, API and MCP calls now return a clear "Try again later" error with a request ID instead of raw internal connection details.

Environments

  • Admin Limit which MCP servers each environment can use. Environment pages now have an MCP servers section where you can turn off every MCP server for that environment, or list the servers its agents may use, one per line as a URL, hostname pattern, or command. Leave the list empty to allow every server your team allows. This is now available to everyone. See Environment management.
  • Each environment gets its own Secrets page and Edit dialogs. The Add secrets dialog warns before replacing an existing secret and refuses duplicate names. On an environment's page, change the name, staleness threshold, and install timeout from Edit dialogs, promote it with a Promote to team button, and delete it from a Danger zone section. Renaming warns when the name is taken, Version history marks the Current version, and duplicating or promoting an environment keeps its settings. See Environment management.

Pull requests

  • Admin Cloud Agents open a pull request by default, including multi-model runs. Cloud Agents now open a pull request for every run unless you turn PR creation off, including runs that sample several models. The Create PRs setting in your personal and team Cloud Agents preferences is now a single Always create PRs switch, and the old For Single Model Runs option now behaves like always on.

MCP

  • MCP permission requests work with more servers, including Cloudflare. When an MCP tool needs an extra permission, Cloud Agents can now ask for any permission the server publishes as available, so tools from servers like Cloudflare prompt you to sign in again instead of failing. Requests that name several permissions are read in full.

Integrations

  • Admin Choose when Cursor responds in its Slack threads. Team admins can now set a Reply mode for the Slack bot in Cloud Agents settings. Follow threads, the default, lets Cursor pick up plain replies in its threads when they're useful, while Mentions only limits it to replies that @mention Cursor or answer a question the agent asked. Mentioning @Cursor with no message in a thread with a running agent now has the agent read the thread and continue, and replies from other bots no longer answer an agent's open question. See Slack.
  • Sign in to MCP servers from Slack. When a Cloud Agent you start from Slack needs you to sign in to an MCP server, it posts a sign-in card in the agent's Slack thread. The card's button opens that server's sign-in directly with a single-use link, and Skip declines the sign-in so the agent can continue without it. See MCP tools.
  • Shorter, first-person agent messages in Slack. Agents started from Slack now keep their replies short and skimmable, and status messages speak in the first person, like "I'm done." or "Got your follow-up." The "Open in Cursor" footer now appears only on the agent's first message in a thread, and Slack no longer sends a welcome tip after your first @Cursor launch. This is now available to everyone.
  • Choose any repo you can access when launching Cloud Agents from Jira. Add repo=owner/repo to name any repository your linked account can reach, even one outside your routing list, and get a clear message if it is not found or you lack access. When you assign an issue to Cursor, a repo= in the issue description picks the repository unless your trigger text names one. Bare workspace/repo names on Bitbucket Cloud now resolve to Bitbucket instead of GitHub, and users who still need to connect their Cursor account are prompted again instead of seeing an error. See Repository selection.

Self-Hosted Machines

  • Admin Team pools can receive source control tokens and dashboard secrets. Self-hosted pool workers can now get short-lived source control tokens and your team and user Cloud Agent secrets during runs. A team admin turns on SCM token minting or Secret sync in the Self-Hosted Machines section of the Cloud Agents dashboard, and each worker opts in with --mint-scm-token or --sync-dashboard-secrets. See Environments on Self-Hosted Machines.

Network and security

  • Agents can check which network hosts they can reach. The built-in Cursor Cloud MCP adds a network-info tool that reports the network policy enforced on the agent's machine: unrestricted, or the exact allowed domains and TCP destinations. Agents use it to explain a blocked request instead of probing hosts one at a time, and it flags when a just-changed policy may not have applied yet. environment-info now reports the same details.

Dashboard

  • Admin Repository routing rules get their own panel in Cloud Agents preferences. Add, edit, and delete the rules that pick a repository or environment there, and open Slack settings from a single row that links to the Slack page. The default model is now chosen from one menu with search and model options, Base branch and Branch prefix show the default that applies and change through an Edit dialog, and you can pick Origin repositories shared with you as your personal default repository. See Default settings.

Deployed agents

  • Admin Per-agent roles now control deployed agents for every team. Deploying, stopping, deleting, or editing secrets of a deployed agent, one listed on the dashboard's Deployed Agents page, now requires the matching role on that agent. Owners can manage secrets, while Editors and Viewers can't. Searching Deployed Agents or Debug agent runs now reads further pages for you and shows Searching until it's done. With Privacy Mode (Legacy) on, Deployed Agents pages show the privacy notice instead of forms that would fail.

Fixes

  • Fixes to agent turns and stopping. Stopping an agent while it's still starting now ends the turn cleanly. Agents no longer get stuck when a follow-up arrives during a failure. Forked agents keep the contents of copied steps, and diffs load for branches the git host hides from ref listings.
  • Fixes to subagents. Stopping a Cloud Agent no longer leaves its subagents stuck as running or drops results they had already returned, including from nested subagents. Subagent calls now appear in the chat as soon as the agent issues them, and an agent waiting on a background subagent shows as waiting instead of idle.
  • Fixes to MCP servers. MCP servers passed when starting an agent from the API or SDK are now available from the first step. A rejected Client Secret now tells you to update it in the connector settings, and MCP servers reached over your tailnet no longer show another team's cached tools or get blocked by Cursor's shared network egress.
  • Fixes to Slack. Agents in Slack no longer repost an earlier reply or start an extra turn when you answer their question. Finished-agent DMs with long summaries now arrive reliably, and Slack install links stay valid for 30 minutes.
  • Fixes to repo skills in the slash menu. Repo skills in symlinked folders now show up in the / menu, repo commands still show on branches not yet pushed to GitHub by using the default branch's commands, and large repos no longer drop skills found before a listing limit. Typing a name now runs the skill or command the menu shows, including skills from your Agent Store. Repos you can't read are explained clearly, for example when you need to connect GitHub.
  • Fixes to Projects. Tracked tasks now pick up merged or closed pull requests from GitLab and Bitbucket, link Azure DevOps pull requests, and no longer close a routine's task when a linked pull request merges. A cancelled routine chat that never started now shows its error instead of looking stuck, and adopted workers without a name show as "Untitled chat" instead of a raw ID.
  • Fixes to Self-Hosted Machines. Agents that start on a cold self-hosted pool now get computer use when the worker they land on supports it, and agents no longer get stuck showing as running after their turn ends. Workers registering at the same time no longer fail intermittently, brief discovery outages are less likely to stop agents from starting, and service account keys with a repository scope work for worker requests.
  • Fixes to repository access and pull requests. Agents on Bitbucket Cloud now recover from expired credentials, and a failed repository access check names each repository and the reason. Draft pull requests stay drafts, pull request links in multi-repo agents point at the right repository, and a pull request's status no longer gets stuck as not found.
  • Fixes to environments. Saving an environment setup no longer fails with a "Draft environment not found" error. The Environments page keeps your environments listed with a retry option when the source control check fails and no longer flashes controls while loading. Update stale builds now rejects anything but a whole number of hours.
  • Clearer errors when an agent can't start or continue. A brief failure to read your privacy setting now shows as a temporary error instead of a Privacy Mode block. Follow-ups blocked by team settings explain why and keep your message, and Linear-launched agents and best-of-N judging work when your team's model policy blocks the default helper model.
  • Fixes to app publishing. Republishing an app from a Cloud Agent with one repository now uses the folder the app was published from. The security review before publishing no longer flags its own notices as prompt injection on large apps, and making a published app public no longer requires the agent that first published it. Saving the secrets an app asked for now briefly shows Saved before the card closes.
  • Fixes to the Cloud Agents dashboard and web app. Cloud Agent pages now recover on their own when a page load is cut off. Default model pickers in settings no longer jump in width while loading, secret fields no longer get filled with saved passwords, and script fields no longer autocorrect. Saving quick actions keeps the ones from your other teams.

Integrations

Bitbucket

  • Bugbot and @cursor work on every repository in a linked Bitbucket workspace. When the Cursor app is linked to your team, Cursor now handles pull request events from any repository in that Bitbucket Cloud workspace, including new repositories and ones nobody has opened in the Cursor dashboard. Before, those events were ignored. Picking Cursor from Bitbucket's mention autocomplete in a pull request comment now runs the command the same as typing @cursor.
  • Fixes to Bitbucket Cloud connections. Bitbucket Cloud accounts no longer get disconnected when a token refresh coincides with a failed account lookup. The connection stays linked instead of showing as not connected.

Slack

  • Admin Slack settings in one place. A new Slack page under Integrations in the dashboard lets you connect, reconnect, or unlink your Slack workspace, turn Slack notifications on or off, and pick which team runs agents started from Slack. It links to your routing rules, and the Manage rules button and rule links in Slack now open the right rule in the right team or personal workspace.
  • Clearer Slack errors when a cloud agent can't verify repository access. When a cloud agent started from Slack fails because Cursor can't verify access to a repository in its environment, the failure message now names each affected repository and the reason, such as missing write access, missing read access, or a Git remote that is rewritten to another URL.

GitLab

  • Admin Cloud agents only get GitLab write access to submodules and dependencies you can push to. When a cloud agent works on a GitLab project, it now gets credentials for the project's submodules and dependency repositories only where you have Developer access or above, or are an instance admin. Repositories you can't push to are left off the agent, and Cursor creates no project access token for them.
  • Admin Sync GitLab repositories with a team API key. A team API key with Cloud Agents access can now start the same GitLab repository sync as Sync Repos in the dashboard, so automations pick up new GitLab projects without anyone opening the dashboard. The sync runs as the admin who created the key. If that admin is no longer a team admin or hasn't connected the GitLab instance, the request fails with an error that says what to fix. Syncs are rate limited per team and recorded in the team audit log.

GitHub

  • Admin Fixes to GitHub repository access and GitHub Enterprise Server connections. Team admins who haven't connected a personal GitHub account now see the team's GitHub organizations and repositories instead of a prompt to connect GitHub. GitHub Enterprise Server connections no longer break when the server redirects to a URL with a different scheme.

Integrations page

  • Fixes to the Integrations page. Source-control cards such as GitLab Self-Hosted now load at their final size instead of growing when their status arrives. Menus show the full reason an action is disabled, and buttons that open a new tab show an outward arrow.

Dashboard and admin

API keys

  • Admin See every kind of API key in one list. The API keys page now shows all your keys in one table with a Type column, and admins who manage keys see team and personal keys together with each key's owner. Create lets you pick the key type in one dialog, with a Team key switch to pick who owns the new key, and a Filter menu narrows the list by type and status.
  • Admin Find keys created by people who left your team. Admins who manage API keys now see a Created by people no longer on this team section on the API keys page. It lists the API keys and service accounts that former members created, so you can revoke the ones you no longer need.
  • Admin Give team API keys access to automations. When you create a team API key, the permission picker now offers View automations and Manage automations. View lets a key read the team's shared automations, including their prompts, triggers, and repository settings, and Manage lets it change shared automations the team allows others to edit. Neither covers members' private automations.
  • Admin SSH keys moved to My settings and Origin. Your personal SSH keys now live in the SSH and signing keys section of My settings. Team SSH certificate authorities and the IP allowlist moved to the Security section of the Origin page, and old links to the SSH tab on the API keys page redirect.

Security and audit logs

  • Admin Edit each Network access list in its own dialog. In workspace settings, Network access now shows the sandbox allowlist and blocked destinations as rows with a destination count, each with its own edit dialog. The Cloud agent egress uses the sandbox allowlist switch is on by default; turn it off to give cloud agents their own allowlist row. Saving other team settings no longer overwrites your network lists. See network access settings.
  • Admin Clearer audit logs, and a record of a removed member's keys. When you remove a member, the credentials revoked event in your audit log stream now lists the team API keys, fine-grained keys, and service accounts they created that still work. The audit log table shows separate Date and Time columns with sentence-case event names, and you can filter for Origin namespace access policy events. See audit logs.
  • Admin Protect an organization from one dialog. Admins now protect an organization or group from a single Protect organization dialog that takes its URL. With nothing protected yet, the card shows a Protect organization button, or a prompt to connect GitHub or GitLab first. See Protected Git Scopes.

Plugins and MCP

  • Admin Limit which MCP servers each cloud agent environment can use. A cloud agent environment's page now has an MCP servers section where you list the allowed servers one per line as a URL, hostname pattern, or command, or turn on Disable all MCP servers. Agents in that environment then get only those servers from the ones your team allows. This is now available to everyone.
  • Admin Set a team plugin's distribution with switches. A team plugin's details dialog on the Plugins page now has Required for the team and On by default switches that apply when you click Save. In the plugin variables dialog, leaving a saved secret blank keeps its value, and Clear variables removes them.
  • Admin Import a team marketplace from a specific branch. When importing a team plugin marketplace from a repository, admins can now enter a Branch instead of using the repository's default branch. The Manage marketplace dialog also shows the reason when a save or delete is refused. See adding a team marketplace.

Cloud Agents

  • Admin Repository routing rules moved to Cloud Agents preferences. Manage the rules that pick a repository or environment for requests from Slack, Jira, Linear, or the API in a Routing rules panel on the Cloud Agents preferences page, for your team or your personal workspace. Older routing links and Manage rules in the Slack integration open that panel with the linked rule highlighted. See routing rules.
  • Admin Cloud Agents PR creation is an on/off switch. The Create PRs dropdown in Cloud Agents preferences is now an Always create PRs switch in team and personal settings. Teams that had "for single model runs" selected now get a pull request for every run, and in personal settings, leaving it unset follows your team's choice.
  • Pick a default model and its options in one menu. The Default model setting in Cloud Agents preferences is now a single picker: search for a model, then set options like effort or fast mode in the same menu.
  • Admin Run Approval Agents when a pull request joins a stack. Approval Agent triggers on GitHub-only repositories now include Pull request stacked, which fires when a pull request is added to a GitHub stack, drafts included. See configuring triggers.

Usage, members, and groups

  • Admin See which products used your weekly included usage. The weekly usage meter on your plan card and the Usage page now splits into colored segments by product, with a legend showing each product's share. The Usage chart also shows Cloud Agent spend under its own label instead of Other.
  • Admin Break down usage charts by member. The daily usage chart on team and organization Usage pages has a User option in its group-by menu, so admins who can see members can break spend or tokens down by member.
  • Admin Spend alerts require a recipient. Creating or editing a spend alert now requires at least one recipient, keeps at least one notification channel selected, and lets you search long admin lists.
  • Admin Paste invite lists from your email client. In the Invite members dialog, you can paste addresses in the Name <address> format from an email To line, and Cursor invites only the addresses and flags broken entries. An expired, revoked, or unknown invite link now says "This invite link isn't valid", and an invite sent to a different email than your account explains that and offers to switch accounts.
  • Admin Confirm before changing an admin to a member. On the Members page, changing a team admin's role to Member now opens a confirmation that explains they'll stay on the team but can no longer manage its members, invites, billing, or settings. See changing roles.
  • Admin Group model access saves exactly what you check. Saving a group's model access now keeps exactly the models you checked, and a group's Models page shows whether access is on, with a Turn on or Turn off button that confirms before members fall back to the team policy. You can't delete a group until its members are removed, and the dialog says how many remain. See group model access.

Personal settings and navigation

  • Your dashboard theme follows your account. Choosing Light, Dark, or System on the dashboard now saves to your account, so the same mode appears in every browser you sign in from.
  • Manage cookie preferences from My settings. A new Cookies section under Privacy & security in My settings lets you turn optional cookie categories on or off, or use Accept all or Reject all. Necessary cookies stay on.
  • Settings search shows the exact setting. Searching settings now keeps only the matching rows and highlights the matched words. The Cmd+K command palette also finds individual settings in My settings, such as Theme or Home team, and opens the page scrolled to that setting.
  • Resize the dashboard sidebar. Drag the sidebar's edge to make it wider or narrower, or past its minimum width to collapse it. The width is remembered across visits, and the handle works with the keyboard.

Fixes

  • Admin Saving settings no longer undoes other settings. Saving one section of team settings no longer resets unrelated admin settings such as network allow and deny lists, model lists, or browser origins, and saving MCP allowlist rules no longer flips the Enable Team MCP Allowlist switch. Saving quick actions in one team no longer removes the ones in your other teams.
  • Admin Fixes to usage and members. Usage alerts no longer arrive twice for the same threshold, and daily usage charts include usage from merged teams. Role, seat, and Over limit filters on the Members page now match across up to 5,000 members, with a warning when a longer roster is cut off.
  • Admin Fixes to API keys, MCP connectors, and Cloud Agents settings. The dialog showing a new or rotated API key secret now stays open until you click Done. MCP connectors whose sign-in keeps failing show a Reconnect button, and Cloud Agents settings load and save for the team you have selected.
  • Fixes and polish. Amount fields no longer misread a decimal comma like 10,50 as 1,050, and the Overview, Integrations, and My settings pages no longer jump around while they load. Links to Cloud Agents, billing, and workspace settings open in your team or personal workspace instead of the organization overview.

Automations

Tools

  • Automations can open pull requests ready for review. When an automation opens a pull request, it can now open it ready for review when your prompt asks for it. Pull requests still open as drafts by default. A subagent working in another repository now opens its pull request on that repository, instead of failing or opening it on the repository of the agent that started it.

Automation runs

  • Test runs now tell you why an automation didn't start. When you test an automation and the run is refused, you now see the reason, such as no prompts configured, missing repository access or environment secrets, or a rate limit with "Too many automation runs right now. Wait a moment and try again." Runs that fail because Cursor couldn't check your privacy mode now show as temporarily unavailable instead of as a privacy setting problem.
  • Saving an automation now catches repositories it can't use. When you add a repository, or an environment with repositories, to a cloud automation, Cursor checks that you can access each new one before it saves. Saves with a repository you can't reach or an unsupported repository URL fail with a clear message, so you don't end up with an automation that fails on every run. Repositories already on the automation aren't checked again. See Repositories.

Automations page

  • Retry failed follow-ups in the automation builder. When you describe an automation in plain language and a follow-up message fails, the automation builder now shows the error above the Follow up input with a Retry button. Retry resends your last message, and the draft you already have stays in the form.

Integrations

  • Project automations skip the run-finished reaction in Slack. Slack-triggered automations that a Project sets up no longer add the default run-finished reaction to the triggering message. This applies to new ones, and to existing ones once their instructions, trigger, or destination change. A custom completion emoji still applies.

Fixes

  • Template repositories stay selected while repositories load. When you create an automation from a template on the web, the template's repository is now filled in right away and stays selected while your repository list is still loading. Before, it was left empty if the template opened before your repositories finished loading.

Security and approval agents

Security agents

  • Vulnerability findings show their OWASP Top 10 category. Each finding from a Vulnerability Scanner is tagged with its OWASP Top 10:2025 category, so you can see the class of issue at a glance. The category appears as a badge on findings in the dashboard and in Slack finding replies, and the new OWASP category filter in Reviewing findings narrows the list to one category.
  • Vulnerability findings lead with their impact. When you open a vulnerability finding, its impact appears in a highlighted callout at the top, so you see why it matters first. Reports also have clearer descriptions and labeled evidence, and the scanner rates findings high or critical only when they're tied to a concrete business consequence.
  • Fixed docs links on security dashboard pages. The Learn more link on Security Review and Vulnerability Remediation pages, including each agent's page, opens the docs for that kind of agent instead of a generic page. When a scan fails because the default branch has no commits, the notice links to the Vulnerability Remediation docs.

Approval agents

  • Admin Name your approval policies. On Origin repositories, a PR Routing & Approval policy can set an optional name, such as Public API contract, so two rules with the same owners no longer read the same. A name is one line of up to 80 characters, unique within its policy file, and titles the rule on the pull request page, in origin pr requirements, and in approval comments. Policies without a name look the same as before.

Bitbucket

  • Bugbot works on Bitbucket repositories added after setup. Fixed an issue where Bugbot ignored pull requests and comments on Bitbucket Cloud repositories added to a connected workspace after setup. Bugbot now picks them up automatically.
  • Fixed Cursor mentions picked from Bitbucket autocomplete. Choosing the Cursor app from Bitbucket's mention list now runs your comment command, the same as typing @cursor.

Fixes

  • Fixes to review reliability. Bugbot on Bitbucket no longer misses pull request updates when Bitbucket rate-limits it. It waits for the limit to clear and retries. A review that already posted its findings now keeps them, and its completed status, when a later retry is canceled.

Agents Window

Projects

  • Project coordinators offer only the modes they support. In a Project's coordinator chat, the Modes row, slash commands, and Shift+Tab cycling now offer only Plan, leaving out Ask and Debug, which are meant for single-agent work. If you already picked Ask or Debug for a coordinator, its mode pill stays until you dismiss it.

Chat and composer

  • Drafts are no longer sent twice from different windows. If another window already sent or discarded a draft, sending it again now shows a notice instead of starting a second agent. If another window already sent the same New Chat prompt, you're asked to press Enter again to send it anyway.
  • Images from chat open in an animated full-window viewer. Opening an image from an agent chat now shows it in a viewer that animates open and closed and fills the window with a margin. Closing the viewer returns focus to where you were.

Navigation and layout

  • Keyboard access to the menubar on Windows and Linux. Press Alt+F, Alt+E, Alt+V, or Alt+H to open File, Edit, View, or Help, and hold Alt to see the underlined letters. Inside an open menu, press a row's underlined letter to run it. Settings now sits in the File menu, and About Cursor and the update actions are at the bottom of Help, matching the Cursor IDE.
  • Open with Cursor starts a new agent in the Agents Window. On Windows, right-clicking a folder in File Explorer and choosing Open with Cursor now opens it as a new agent in the Agents Window when that was your last active window, the same as running cursor <folder>. Previously it opened a separate editor window instead.

Customization and settings

  • Zoom now scales code along with the rest of the Agents Window. Zoom In, Zoom Out, and Reset Zoom now scale code editors and diffs too, instead of leaving code at a fixed size. In Settings > Typography, UI Font Size is replaced by a Zoom setting shown as a percent from 50% to 300% in 10% steps, matching the zoom commands, and your previous size carries over. Code Font Size stays independent of zoom.

Files, changes, and pull requests

  • Live PR status in plans. Pull request links in a plan now show a status icon that updates as the PR changes. The PR tab no longer repeats the Open in Cursor, Open in Web, and View Automation badges in pull request descriptions and comments, since the Agents Window already offers those actions.

Cloud agents

  • Cloud subagents point you back to their parent agent. Subagents started by a cloud agent can't take follow-ups. When you open one, the prompt box now explains this and tells you to message the parent agent, and an Open parent agent button takes you straight there.
  • Origin repositories show up in more repository pickers. When you add or edit a secret in environment setup, the Apply to picker now lists your Origin repositories next to your GitHub ones. The Automations repository picker shows Origin repositories before you start typing. When a repository with the same owner and name exists on both GitHub and Origin, the project picker and the secret picker show which host each one is on.

Fixes

  • Fixes to the agent transcript. Editing a sent message no longer makes its text jump, and the cursor lands at the end. Messages only slightly too long now show in full instead of behind Show more, a chat's loading skeleton no longer sticks if you scroll early, and prompt tray hovercards stay beside their row.
  • Fixes to file links in cloud agent transcripts. Clicking a relative file path in a cloud agent's transcript now opens the file from the agent's actual repository checkout, including on self-hosted workers.
  • Fixes to slash commands in cloud agents. A slash command followed by punctuation, like /deploy?, now runs as a command.
  • Fixes to Projects. Filtering the sidebar by Source or unchecking Cloud under Environment no longer hides your Projects. The Projects credit card in the sidebar no longer disappears after you switch accounts, and on a task Board with no visible columns, the list of empty statuses now spans the full width.
  • Fixes to Annotate in the browser. In Annotate, clicking a spot on the page that can't be picked no longer leaves the note card hidden, and empty notes close cleanly. Notes drop their empty header row, show picked elements as badges, and drag only from above or below the text, so clicking beside a line places the cursor. In responsive mode, the CSS inspector docks to the browser corner and you can start a selection beside the device frame.
  • Fixes to keyboard navigation and accessibility. Opening a pull request from the PR stack or the Changes tab now moves focus into the new PR tab, and arrow keys in the Library search box move into the grid. Tab close buttons no longer take keyboard focus, menus show your actual shortcuts for Find in Changes and new terminals, and Escape during IME composition or Shift+Escape no longer closes trays. Screen readers also announce the correct label when the Update button only restarts Cursor.
  • Fixes to chats and the sidebar. Custom icons set with Edit Icon on local chats now stick, a quick reload on a New Chat draft no longer blanks the window, and the model you pick when starting a local agent from Home is kept for the next draft. The conversation minimap no longer breaks the transcript, and revealing deeply nested files in the Files tree is faster.

Cursor IDE

Agent and chat

  • Fixed Escape during voice dictation. Pressing Escape while you dictate in the agent input now stops the recording and keeps your cursor in the input with the dictated words. Previously Escape left the input and the mic kept recording.

Customization and settings

  • Hooks can tell which agent or subagent type spawned a subagent. Hook scripts now receive parent_agent_name alongside parent_agent_id: the subagent type for a subagent, or the custom mode name for any other agent. subagentStart and subagentStop hooks now report the parent turn's generation_id instead of the conversation ID.
  • Fixed Customize opening local files in remote windows. In a remote window, opening or deleting a user skill, rule, or subagent from Customize now acts on the file on the remote host instead of a local path.
  • Fixed skills missing in multi-root workspaces with nested folders. When one folder in a multi-root workspace sits inside another, skills at the inner folder's root are now offered to the agent instead of being hidden.

MCP

  • Admin Set up the Asana plugin without your own OAuth app. When admins set up the Asana plugin, the form no longer asks for a client ID or secret. Cursor now provides the Asana OAuth app.
  • Admin Easier Databricks Genie One plugin setup. When admins set up the Databricks Genie One plugin, the Workspace URL field now accepts any workspace or Genie MCP URL and keeps only the hostname, and an invalid URL gets a clearer error.

Models

  • Admin Private Inference model picker no longer loads forever. With Private Inference on, the model picker stops waiting after about a minute and says models are unavailable instead of showing Loading models indefinitely. If this machine has no Private Inference config, the picker tells you to contact your admin.

Fixes

  • Fixes to agent reliability. Agent edits no longer hang on malformed patches, searches across all files no longer fail, and the agent can read PDFs again. Image generation now says when a reference image is missing.
  • Fixes to agent questions and chat history. Answering an agent's question now always reaches the agent, including after reopening the chat or restarting Cursor, and the question keeps its options. New chats no longer disappear from chat history after the editor crashes during their first turn. Follow-ups, Resume, and Send Now in a subagent's chat no longer duplicate reply text or drop the message.
  • Fixes to MCP servers. Plugin MCP server URLs with a ${VAR} placeholder in the path now expand correctly, and servers whose config changed while connected, such as a rotated auth header, restart cleanly on reload. MCP sign-in now completes in remote windows, and saving cloud agent MCP defaults no longer stores built-in entries.
  • Fixes to editors, updates, and chat tabs. Multi-file diffs no longer leave broken entries or show outdated files, and the first launch after an update no longer shows a stale update banner. Archived chats no longer reopen as editor tabs on restart. On Linux, the updater no longer targets another AppImage when Cursor is launched from its terminal.
  • Fixed pending agent edits rewriting files at startup. When Cursor reopens agent edits you haven't accepted or rejected, it now shows them over the file as it is, without editing or saving the file. Previously this could save outdated contents over a file changed elsewhere, sometimes leaving it empty.

Cursor CLI

Fixes

  • Fixes to Plan mode. Plan mode no longer lets the agent edit a non-markdown file through a symlink with a .md name. Those writes are now refused with the usual Plan mode error.

Cursor mobile

Voice calls

  • Voice calls give up-to-date answers about the agent's work. A voice call now knows the recent chat, including the last week of messages and earlier calls, and keeps up with the agent's branch, pull request, test runs, and to-dos as they change, so asking how the work is going gets a current answer. A request that can't reach the agent within 20 seconds is reported as unreachable instead of being sent late, and a call that starts or reconnects while the agent is idle or already working no longer resends your last request.
  • Voice call answers are shorter and stick to what the agent knows. Answers start with a short spoken summary of about 30 words and offer to go deeper, and calls no longer read out URLs, long file paths, full commit hashes, or tool names. When you ask for an exact value or a code walkthrough, the call looks it up instead of guessing, and it reports test and CI results without speculating about why they passed or failed.
  • Talk over the agent in voice calls. The agent's voice fades down as soon as you start talking and comes back if it was only background noise. When you interrupt a reply, the agent knows which part you heard and can pick up the rest instead of repeating or dropping it.
  • Voice calls read out the agent's plans. When the agent ends a turn with a plan during a voice call, the call reads out the plan's name, overview, and to-dos instead of only saying it needs you in the chat. The full plan stays in the chat.
  • Short tones tell you what the agent is doing. During a voice call, a soft tone plays when a task is sent, when a result is about to be read after a pause, when the agent needs your decision, and when something goes wrong. Tones play alongside speech without interrupting it. Turn them off with Status chimes in the voice call settings.
  • Voice calls understand your repo, file, and agent names. Voice calls, including on CarPlay, recognize spoken repository, branch, file, and agent names from the current conversation more reliably. Asking to send to or switch to an agent works with names in Japanese, Chinese, Korean, Thai, Arabic, and accented Latin, and interrupted replies in those languages pick up at the right spot.
  • Voice calls show on CarPlay. When a voice call plays through your car's CarPlay audio, its Live Activity appears on the CarPlay display with the agent, a live timer, and a Hang up button. Muting from the Live Activity or the system call screen stays in sync, and a Live Activity you dismiss stays dismissed for that call.

Composer

  • Edit your words while you're still dictating. Tap the live caption while dictating in the composer to edit what you've said so far. Your edits stay, and anything you say next is added after them without doubled spaces or punctuation. Dismissing mid-dictation now only hides the keyboard.

Account and billing

  • Clearer warning before deleting an account merged with Grok. If your Cursor account is merged with a Grok account, the Delete Account page warns that deleting it also permanently deletes your Grok account and any Grok plan, with no way to restore them.

Fixes

  • Voice calls. What you say while a call is still connecting is no longer lost, and a two-part question no longer ends in silence. The call no longer greets you after it has already started talking, and it keeps track of the agent's latest progress during long tasks. Voice call cards stay visible and attached to your last typed message.
  • Foldable layouts and large text sizes. Folding or unfolding your phone, or switching iPad layouts, keeps your place in the chat you were reading, and returning from the background no longer resets navigation. At the largest accessibility text sizes, sidebar icons, round buttons, and the composer's repo row and toolbar no longer overlap or clip. Slash command suggestions also come back when you return to a draft.
  • Conversation loading and sign-in. Sign-in and session refresh work even when your phone's clock is wrong. Opening a large Project loads back to your last message instead of stalling, and a follow-up sent before leaving a chat shows in the right place when you return. The paywall sheet no longer shows two back controls.

Cursor CLI

Plan mode

  • Plan mode no longer lets the agent delete or change code files. In Plan mode, the agent can no longer delete non-markdown files, and it can no longer get around the markdown-only rule by writing to a file:// path. While you plan, only markdown and canvas files can change.

Self-hosted workers

  • sessionStart hooks on self-hosted workers know where the agent was launched from. On a self-hosted worker, sessionStart hooks now receive a surface field naming where the Cloud Agent was started, such as desktop, web, mobile, slack, linear, cli, sdk, api, or automations. Hooks can use it to set up each session by its source. The field is left out when the source is unknown.
  • agent worker debug accepts --auth-token-file. agent worker debug now reads a token from --auth-token-file the same way agent worker start does, so you can check a worker with the credentials it runs with. It reports which auth method it used, and errors if you pass both --auth-token and --auth-token-file. See the worker options.

Fixes

  • Fixes to skills and file search. When one workspace folder sits inside another, the agent now lists skills defined in the inner folder up front, not only when it works on files there. When the agent searches an absolute path outside the workspace and finds nothing, it now gets a hint to search that directory directly, so it recovers instead of reporting no matches.

Agents Window

Projects

  • Arrange and tag Project tasks more precisely. In a Project's task List and Board, you can now drag a task to an exact spot within a status, and the Board makes room where it will land. Empty statuses stack as compact rows after the last Board column, where you can still drop cards or change their icon. Tasks show a chip for every tag, the Tags filter matches any selected tag, and untagged tasks appear under None.
  • Filter and search tasks across all your Projects. The Tasks tab that gathers tasks from all your Projects now has the same Filter button and chips as Project notes, plus a Project filter, so you can narrow tasks by Project, Status, PR, Tags, or Owner. Press Cmd+F or click the magnifier to find tasks by text, including Project names, in List or Board view. The tab only lists Projects you belong to.
  • Projects move to tracked tasks. When you open a Project you own that still keeps its tasks in notes.md, the Agents Window now moves them into a tracked task list and the notes tab switches to the tasks board. Headings that hold only notes stay in notes.md.
  • Choose default models for Project coordinators and workers. Settings > Models has a new Projects section where you pick the model a new Project's coordinator starts on and the model its workers use when the coordinator names none. The new Project picker starts on your default coordinator model, and you can still change it per Project.

Chat and composer

  • Voice calls know more and hear you better. Voice calls start already knowing the agent's status, active file, branch, changed files, last command, and todos, so simple questions get answered directly. They better recognize spoken repo, branch, file, and agent names, including accented, CJK, and Thai text. On a call with a Project coordinator, the call speaks the coordinator's reply without a redundant "It's finished". A new request you speak while the agent is working now gets its own answer instead of counting as already answered, and past call cards stay next to the turn where the call happened.
  • Zoom into chat images with a click. In the image viewer, clicking an image now steps zoom from fit to 2.5x to 5x, and Option-click (Alt-click on Windows and Linux) steps back out. Dragging stops at the image edges, and clicking any empty area closes the viewer. The full-screen preview for prompt attachments adds an image counter and a toolbar.
  • Drafts keep attachments across reloads and restarts. An unsent prompt in a cloud agent chat now keeps its attached images, files, and mentions after a window reload or app restart, including image-only drafts. Previously only the text came back.

Files, changes, and pull requests

  • Approve, take over, and fix conflicts from the PR tab. Origin pull requests now show an Approve button in the PR tab header. It switches to Re-approve after new pushes and reads Approved once your approval covers the latest changes. If you have permission to take over a pull request, Take over pull request in the PR tab menu asks for a reason and makes you the author, and Fix Merge Conflicts is in the menu for draft pull requests. When an agent has several PRs with failing CI or conflicts, the Debug CI and Fix Conflicts pills show a count and let you pick which PR to act on.
  • Pull request links open where you expect. Pull request links in the PRs tray, apps rail, Changes tab, and View PR picker now open where your PR Link Destination setting says, and Cmd+click (Ctrl+click on Windows and Linux) opens one the other way. Middle-click does the same in the PRs tray, apps rail, and View PR picker.
  • Add folders to chat, and Option-click files, chips, and tasks to add them. Right-click a folder in the Files tree and choose Add to Chat to insert it as a folder chip, the same one the @ menu adds. Option-click (Alt-click on Windows and Linux) a file in the Files tree, a chip or link in the transcript, a subagent card, or a Project task to add it to chat without opening the menu.
  • Jump around long Markdown files with a table of contents. Wide Markdown previews show the document's headings in the right margin. Click one to scroll to that section, and the section you're reading stays highlighted, even as you edit. Opening a link to a specific line of a Markdown file now keeps the preview and scrolls to that line instead of switching to Source.

Navigation and layout

  • Undo an archive right from its toast. Archiving or unarchiving an agent now shows a toast with an Undo button, and each new archive replaces the previous toast instead of stacking. For a few seconds afterward, Cmd+Z (Ctrl+Z on Windows and Linux) undoes it and Cmd+Shift+Z redoes it, even while a text field has focus.
  • Self-hosted agents show the right machine name. Agents on a self-hosted or Remote Control worker now show that worker's machine name in the sidebar, status bar, and tooltip, instead of labeling another machine's run as this computer.

Cloud agents

  • Clearer secrets in environment setup. Secrets in the Environment setup pane now use a paged table and a new Edit Secret dialog with repository scope. Managing secrets for a personal environment opens My Secrets on the dashboard instead of Team Secrets.
  • Find more of your repos in the project picker. The project picker now finds Origin repositories in every namespace you can read, including team and shared namespaces, and searching for owner/name looks in that owner's namespace. A repo that is only part of a multi-repo sidebar project can now be added as its own project, and multi-repo cloud agents keep each repo's tracked branches in the editor panel's branch picker.
  • See which self-hosted machine a subagent runs on. Subagent cards now show a machine icon when the subagent runs on a self-hosted worker, whether it was sent to one or inherited its parent's machine. Hover the icon to see the machine's name.
  • See automations and deployed agents in one list. The Automations panel now lists your automations alongside your team's deployed agents in a single table, with a 15-day run history chart on each row and options to sort and filter. Deployed agents open on the dashboard when clicked. Environment pickers in Automations now list every environment instead of stopping at the first 200.

Canvases

  • Canvases use a new page layout. Canvases the agent builds now use a page-style layout with sections, diffs, task lists, usage bars, and richer charts, and headline metrics line up in a grid of up to four columns.
  • Canvas sharing shows Project access and clearer access errors. When a Team Project's agent creates a canvas, its Share menu in the Canvas tab now lists the Project as a read-only Viewer. Opening a canvas you can't see now says "You don't have access to this canvas" instead of saying it isn't available.

Fixes

  • Fixes to queued and steered messages. Steering messages sent at the end of a turn now appear in the conversation, subagent follow-ups no longer show text twice or drop the message, and answering an agent's question resumes it even after reopening the chat.
  • Fixes to cloud agents. Cloud agent workspaces now reconnect when the agent moves to a new machine, and you no longer see an error when automatic reconnection stops retrying. Steps that fail to load now show as failed instead of reloading forever, and teammates' names stay on their messages after you reopen a shared agent. If you allow an agent's request to run on a machine after the request has closed, a warning now says your approval arrived too late and how to grant access again. Artifact files from a cloud agent keep opening in the Files tab after the agent's machine shuts down.
  • Fixes to Projects. Leaving or being removed from a Project now hides your chats there and shows a removal notice in every open pane, and rejoining brings them back. Leaving a Project also removes its subagents and workers, and new invites show up as soon as you return to the window. Notes from workers no longer split a Project conversation into extra turns or take over the todo card, and access prompts are shorter and link to the environment. In the @ menu, teammate and Project rows no longer carry a redundant "Project member" or "Project" label.
  • Fixes for actions on agents that haven't loaded yet. Choosing a mode or a model from /model, typing a prompt, saving a plan, closing the plugin menu, using Add to Chat from the PR tab, and running Abort Chat no longer fail when the agent's conversation isn't loaded.
  • Fixes to the subagent and question trays. Local subagents that already stopped no longer show as running, and running subagents no longer show as stopped by mistake. A cloud agent's open question now appears in the question tray without expanding its collapsed work group, and pull request links in questions show as PR chips.
  • Fixes to the Agents Window interface. Pressing Escape while dragging now only cancels the drag, and opening the Apps panel in a narrow or maximized window no longer squeezes or overlaps the chat. The composer no longer jumps or overflows when a side panel opens or closes, and the changed-files list keeps updating during heavy file activity. Screen readers now announce alerts and status messages, including while a dialog is open.
  • Fixes to dialogs. Pressing Escape with stacked dialogs now closes only the top one, and closing dialogs like Clone Repository or New File returns focus to the control that opened them. Footer buttons stack instead of overflowing in narrow windows, and Escape in the SSH and WSL connect dialog goes back a step instead of closing it.
  • Fixes to the conversation minimap. The conversation minimap no longer freezes the app when its tab is in the background, and no longer flickers when you resize the pane near its width limit. The hover highlight on its ticks now follows the pointer smoothly instead of stalling.
  • Fixes to hooks and Linux update notifications. Hooks that a local subagent runs before its chat finishes loading now receive parent_agent_id. On Linux installs from the deb package, the update notification shows the same version number as About instead of the package's longer version string.

Cursor IDE

Editor and app

  • Cmd+K now adds your selection to Agent. Inline edit and the terminal Cmd+K prompt bar are retired. Cmd+K (Ctrl+K on Windows and Linux) in the editor or terminal, and Cmd+Shift+K (Ctrl+Shift+K) in the editor, now add the selection to your current agent chat, the same as Cmd+L (Ctrl+L), and keys you bound to the old commands do the same. Resolving a merge conflict now opens an agent chat for that conflict.
  • Update installs in Cursor, even from out-of-date versions. The Update button on the out-of-date banner and update prompt now installs the new version in Cursor, shows progress, and then offers Restart to Update. It opens the downloads page only if it can't update. On macOS, Restart to Update no longer gets stuck on Installing when background activity is turned off.
  • Security fixes. Relaunching a terminal in an untrusted workspace is now refused, and an untrusted workspace's settings can no longer widen terminal.integrated.allowedLinkSchemes. Notification messages show plain text literally instead of parsing it into links, and webviews check file paths on remote machines more strictly.

Agent and chat

  • Wider agent chat by default. Agent chat content now grows up to 1200px wide by default, up from 840px, so code blocks and long responses use more of a wide panel. Set your own limit with the cursor.chatMaxWidth setting.
  • Copied tables paste into spreadsheets and Slack as real tables. Copy Table in the right-click menu of an agent response table now pastes as rows and columns in spreadsheets and Slack, instead of as Markdown text. Choose the new Copy as Markdown item in the same menu when you want the Markdown source.
  • Fixes for non-English text and emoji. Cursor Tab suggestions, agent output, and MCP server logs no longer show garbled characters or broken half-emoji, including Python MCP servers on Windows. Agent PowerShell commands on Windows no longer garble non-English output, including web responses from Invoke-WebRequest or Invoke-RestMethod in PowerShell 5.1. The agent now reads text from most Chinese, Japanese, and Korean PDFs, and image prompts in those scripts are no longer rejected as too short.

Accessibility

  • Screen readers now follow agents and Cursor Tab. Screen readers announce agent turns, replies, errors, and approval requests. In the agent chat, press Alt+F2 to open the latest reply in the Accessible View, or Alt+F1 for accessibility help. Cursor Tab suggestions are announced and can be read and accepted in the Accessible View, and chat buttons and inline diff Keep and Undo buttons can now be reached with Tab and have screen reader labels.

MCP

  • MCP tools ask you to grant a missing permission. When an MCP server refuses a tool call because your sign-in lacks a permission, Cursor now starts a sign-in that adds it, and the agent names the permission the tool needs. If your last sign-in didn't get that permission because you declined it or an admin withholds it, the agent says so instead of returning a generic error.
  • Connect the Box MCP server without your own OAuth app. Adding Box's hosted MCP server now signs you in through Cursor's official Box app in your browser, so you no longer need to register your own OAuth client. If you set your own client ID and secret in mcp.json, sign-in now completes locally with your app instead.
  • Admin Callback URL shown when adding cloud MCP plugins to automations with your own OAuth app. When you add a cloud MCP plugin that needs your own OAuth client ID to an automation in the Automations tab, the setup form now shows the callback URL to register in the vendor's OAuth app. Salesforce plugins also list the scopes to grant.
  • MCP App widgets in chat keep their place after a reload. After a window reload, MCP App widgets in chat reappear at their previous height with a "Loading app" line instead of disappearing. If the server needs authentication, can't connect, or the app fails to load, a compact status row explains why, with Retry for load errors.

Models

  • Search in every model picker. The model pickers in Models settings, the subagent editor, the Automations editor, and when editing a previous message now open as a searchable menu, so you can type to find a model instead of scrolling.
  • Admin Subagents fall back cleanly when their model can't run. A custom subagent pinned to a model your team blocks, that needs Max Mode, or that is unavailable now runs on the parent agent's model instead of failing, and the agent is told which model ran. Subagents on Composer 2.5 Fast no longer fail with an Unknown model ID error.
  • Admin Private Inference fixes for local models and bug reports. With Private Inference on a local provider, agent turns and inference now use the model catalog discovered for the model picker, so local models are no longer misrouted after an extension host restart. Teams whose admin allows bug reports can now send them from Report Issue instead of having them blocked.

Customization and settings

  • Agent follows nested AGENTS.md and rules when it edits files. When the agent edits or creates a file in a subdirectory, it now receives that directory's nested AGENTS.md and always-apply rules, even if it never read a file there first. Each rule is added once per conversation.

Fixes

  • Fixes to agent tool calls. Finished tool calls in chat no longer flip back to in-progress, show as completed promptly, and subagent calls render as subagent cards instead of empty rows. The todo list no longer shows stale or duplicate items.
  • Fixes to restoring agent chats. Chats with corrupted or missing saved messages or checkpoints now open with a fallback for the damaged part instead of failing to load, and earlier turns no longer appear twice. A chat locked by corrupted checkpoints unlocks once you revert past them, restoring a checkpoint after moving a chat to the cloud keeps its embedded context, and archiving or reopening old chats works again.
  • Admin Fixes to MCP servers. Turning MCP servers off and on now reliably hides or restores their tools, the agent gets a server's tools back once its sign-in recovers, and simultaneous OAuth token refreshes no longer discard each other's new tokens. Always Allow no longer saves to the local allowlist when your team manages it, and sandbox and Salesforce sign-in errors now say what to fix.
  • Fixes to Auto-review. Auto-review now sees an MCP tool's input schema even when the server sends it as a JSON string. A beforeShellExecution hook that asks for approval now shows its prompt even when Auto-review would have allowed the command.
  • Fixes to agent chat. Switching between chat tabs keeps each chat's scroll position, and closing a prompt row's right-click menu returns focus to the input. The × on removable pills in the prompt no longer misses clicks, and plugins in the slash menu show their own logos instead of their publisher's.
  • Fixes to the editor and terminal. Running Terminal: Clear while a full-screen program like vim or less is open no longer breaks its display. Inlay hints near the cursor no longer lay out incorrectly or get an unneeded ellipsis, and breadcrumb review controls no longer shift neighboring buttons when you click into an editor group. Renaming a chat tab no longer shrinks it, and the notifications Accessible View no longer jumps after it is dismissed.
  • Fixes to app reliability and settings. Agent browser actions no longer time out on screenshots while the window is minimized. The Remote Control approval prompt keeps digits in device names, and agent startup failures now point to Network Diagnostics in Cursor Settings.

Security

  • The credential helper sends your Origin token only to Origin hosts. The Origin credential helper now gives your Origin token only to Origin hosts. For any other host it declines and says why, so a broad credential.helper setting or a redirect can't leak the token. origin update installs only from Cursor's release channel manifest, so the --manifest-url flag and manifest URL environment overrides are gone. CURSOR_ORIGIN_ENDPOINT must also point to an Origin host over https or to a loopback URL.
  • origin agent setup won't write through symlinks in your project. When you install a project-level skill, origin agent setup now stops if any folder or file on the way to the skill is a symlink, so a checkout can't redirect the write somewhere else. It names the symlink and tells you to remove it or save the output of origin agent setup --print yourself. Global installs under your home directory are unchanged.

Configuration

  • The config directory follows XDG_CONFIG_HOME. When XDG_CONFIG_HOME is set to an absolute path, the Origin CLI keeps your login, settings, and caches in $XDG_CONFIG_HOME/origin-cli. If you already have ~/.config/origin-cli, the CLI keeps using it, so existing logins keep working.

Fixes

  • Fixed clone downloads. origin repo clone now retries brief network interruptions during fast downloads instead of failing. On macOS, it uses parallel range downloads instead of falling back to a slower path.

Web

  • Admin App install screens show who published the app and where you'll be sent. When installing or confirming an Origin app sends you on to the app's own site, the screen now names the namespace that published it, notes when Cursor hasn't verified that publisher, and shows the address you'll be sent to. If you open the create-app form from a link that prefills it, Origin lists each value the link changed and the address each endpoint sends data to, so you can check them before you create the app. Repositories already selected for an app also load right away when you install or edit it, even in namespaces with thousands of repositories.
  • Admin Edit an app's egress IP ranges from the app's page. App managers can now add, edit, and remove an app's egress IP ranges right in the Egress IP Addresses section of its page, each change saved on its own. Origin no longer emails the admins of the app's team when those ranges change. You can also now read and set whether your inbound IP allowlist trusts installed apps' declared ranges through the Origin API; see the Origin API changelog.
  • Admin IP allowlist settings explain what the list covers. The IP allowlist settings now say the list covers git, the Origin API, MCP, and downloads, that Cursor's own services, such as cloud agents, aren't checked, and that once it's on, Origin refuses requests from addresses outside the enabled entries. When the list blocks you, Codebase pages now tell you to connect from an allowed network or ask a Codebase admin to add your IP address.
  • Admin The audit log names who or what made each change. Your team's audit log now names the actor for entries that used to show a generic system actor or nothing at all: app installation tokens show the app, service accounts acting as themselves show the account, and service tokens and team or organization API keys show the key's name. Actions a cloud agent took for you name the agent run, and Origin's own background work, such as mirror sync and scheduled jobs, appears under named Origin system apps. Repository pull entries now include background autofetches from Cursor clients too.
  • Deleting an unmerged base branch closes the pull requests stacked on it. When you delete a branch whose own pull request never merged, Origin now closes the open and draft pull requests based on it, as GitHub does. The pull request's Activity tab says it was closed because its base branch was deleted and names that branch, and a pull request closed because its head branch was deleted now names that branch too.
  • Admin Rulesets can hold more bypass actors and rules. In a repository's Settings > Rules and Protections, a ruleset can now list up to 100 bypass actors and 50 rules, up from 15 and 20. The new limits apply in the ruleset editor, the Origin API, and the Terraform provider.
  • Force pushes now show in a pull request's Activity tab. When someone force-pushes a pull request's branch, for example after a rebase, the Activity tab now shows who pushed and the commits before and after. Pushes without a known pusher are shown as a system event.
  • Cloud agent screenshots and recordings in pull requests now show for everyone who can read the repository. When a cloud agent adds screenshots or screen recordings to a pull request description or comment, Origin now saves them as pull request attachments, so anyone with read access to the repository can see them, even without access to the agent. Images an agent saved to its agent store also load on pull request pages now instead of failing.
  • See which pull requests have auto-merge on from the Pull requests list. Rows in a repository's Pull requests list now show an Auto-merge enabled badge for open and draft pull requests with auto-merge turned on.
  • Find people by @handle in pull request search. In the search box on a repository's pull requests list, the author: and assignee: qualifiers now accept handles with or without the @, so author:@ada works and author:@ lists you and every collaborator. me now means you, the same as @me, and the Me suggestion shows a person icon like the other people suggestions.
  • Admin Find any group when granting repository access. Searching in the Grant Access dialog on a repository's Access tab or the codebase Access tab now finds any team or org group by name, including newly synced groups that used to sit past the first 200 and never appeared. Groups you check stay selected when you change your search, so every checked group gets access when you confirm.
  • Clearer subjects on Origin access emails. When a team or group change gives you access to repositories or codebases, the email subject now names the one resource, the team or group, or the first resource plus a count, for example "You now have access to 7 Origin repos from team Acme." Origin access, app install, and repository share invite emails now come from a Cursor no-reply address, and access emails tell you to reach out to your team admin if a change looks wrong.

API

  • Agents and scripts can upload images and videos to pull requests. The new upload_pull_request_attachment MCP tool and a matching Origin API endpoint upload a screenshot or recording and return a URL you can embed in a pull request description, review, or comment. Through the API, images can be up to 10MB and videos up to 100MB. The MCP tool takes files up to about 750KiB. Only readers of the repository can open the link. get_pull_request_attachment also now fetches screenshots that cloud agents embed in pull request descriptions.
  • Webhooks for repository renames. Origin apps can now subscribe to repository.renamed, sent when a repository's name changes, including a change of case only. The payload includes the renamed repository, its previous name, who renamed it when known, and when, and the event is listed under Repository in the app webhook picker. See the Origin API changelog for payload details.

Fixes

  • Codebase stays in the sidebar when you open a repository you can't read. If you don't belong to a team, opening a link to a repository you can't read no longer hides Codebase from the sidebar.
  • Admin Fixes to permission and token enforcement. Suspending an app installation now revokes its live tokens right away, and webhooks stop for repositories an installation no longer covers, even if already queued. IP allowlists apply to cached app installation requests, deleting a repository honors namespace grants that cover it, built-in roles are no longer confused with custom policies of the same name, and restricted sessions moving branches through the API follow push rulesets. Commit signature verification picks up verified email changes within about ten minutes instead of up to six hours.
  • Fixes to branch, ref, and archive handling. File requests that name a branch like main now return that branch even when a tag has the same name, and archive downloads always contain the exact commit you asked for. Commits and merges that target an invalid ref name now fail right away with a clear error. Committing a web edit to a new branch no longer deletes that branch when the commit may have landed.
  • Fixes to namespace names and app icons. A namespace name shorter than 3 characters now says it is too short instead of just unavailable. App icons and custom emoji now clearly reject SVG files and ask for a PNG, JPEG, or WebP image.

Network and security

  • Admin Customer PrivateLink now covers every PrivateLink destination and works alongside fixed egress IPs. For every team on Customer PrivateLink, Cloud Agents can now reach all of your PrivateLink destinations, such as internal package registries, not only the repo's git host. Teams that also use fixed egress IPs get both: PrivateLink hosts go through your private connection and other allowed traffic leaves from the stable egress range. Git also no longer fails with a CONNECT 403 error under Restricted network access.
  • Admin Revoked API keys and service accounts stop their running agents. When an API key is revoked or a service account is archived or deleted, Cloud Agents running on it now stop within seconds instead of continuing until the next periodic credential check. Agents started by a team API key also stop when that key expires or loses agent access, with a message that names the cause.
  • Admin Clearer notices when your team's inference hook blocks an agent. When an inference hook stops a Cloud Agent, Slack, Teams, Linear, Jira, and pull request threads now show a policy-block notice with next steps instead of a generic error. The hook's reason appears only in private places like the web agent view or a direct message. Your hook endpoint now receives the full tool output, even when that output quotes the refusal message.

Environments

  • Single-repo environments use the same layout as multi-repo ones. Cloud Agents now check out a single-repo environment's repository at /agent/repos/<repo> instead of /workspace, the same place multi-repo environments use, so setup scripts and paths work the same either way. Existing single-repo environments switch on their next build. On the web, file links in the transcript open in the right repository.
  • Environment builds show how long they took and which config they used. The environment builds list in the dashboard now has a Duration column for each finished build. Builds that fail during the Docker build or are still running now show which environment config version they were built from.
  • Fixes for Cloud Agents on Origin repositories. Cloud Agents now find your saved environment and repository secrets for Origin repositories, whichever clone URL you used, and multi-repository agents clone that repository once at the right starting commit.
  • Agents keep starting after you delete an environment. If you start an agent with a remembered environment that has since been deleted, Cloud Agents now runs it on your live environment for the same repositories instead of failing, and worker agents started by a manager follow it to the replacement.

Pull requests

  • Screenshots and recordings show inline in agent pull requests on Origin. When a Cloud Agent attaches screenshots or screen recordings to an Origin pull request, they now upload as pull request attachments, so images display and videos play right on the pull request page instead of appearing as links only people with access to the agent could open. If a file can't be uploaded, the agent is told which one, and images in agent comments on Origin pull requests embed too.

Self-Hosted Machines

  • Admin Project machines and pools follow your team's Self-Hosted Machines settings. If your team has private workers turned off in its Self-Hosted Machines settings, you can't add a pool to a Project's Allowed machines, and if Remote Control is off too, you can't add a machine either. The Machines menu disables those rows with a tooltip saying a team admin has to turn the setting on. Project agents that ask to use such a machine or pool are now refused right away with a message that the setting is off for your team, so nobody gets asked to approve a run that would fail.

MCP

  • MCP sign-in cards come back after a day. When an MCP server refuses a Cloud Agent's tool call for a permission it doesn't name, the agent now offers the sign-in card for that connection again 24 hours after the last offer, instead of only the first time.
  • Agents can't trash other people's Google Drive files. Agents using the Google Drive MCP can no longer trash a file that someone else owns or that lives in a shared drive. They ask you to trash it yourself in Google Drive instead.

Agents

  • Admin Pull request review agents started by service accounts are now read-only for teammates. When a service account launches a Cloud Agent to review a pull request, teammates can still open and read it, but only the service account can send follow-ups, pause it, or act on its pull request, so no one can steer it with the service account's access.
  • Clearer errors when forking an agent whose machine is gone. Forking an agent whose machine no longer exists is now refused up front instead of failing later. The error tells you how to recover: send the original agent a message so it gets a new machine and fork again, or start a new agent if it was archived.

Dashboard

  • Dates on the Cloud Agents dashboard now follow your locale and time zone. Run summaries, run details, charts, and secrets on the Cloud Agents dashboard now show dates in your region's format and your time zone. Previously some of these always used US formatting or the server's time zone, so one page could mix date styles.
  • Open the same agent in the iOS app from Safari. On iPhone, tapping Open on the Cursor app banner in Safari now opens the agent or agents inbox you were viewing in Cursor for iOS, instead of the app's home screen. On the web, Get Cursor for iOS in the account menu now goes straight to the App Store.

Fixes

  • Fixes to stopping, steering, and resuming Cloud Agents. Stopping an agent no longer lets a queued goal continuation start afterward, and stop and steer requests now retry when the backend is briefly unavailable. Answering a prompt that paused a subagent no longer wipes the parent conversation, very long past conversations load reliably, and repositories' git hooks keep working on macOS Self-Hosted Machines.
  • Fixes to dashboard loading and error states. Environment and preferences pages now show an error with a Try again button when they fail to load, instead of claiming the environment wasn't found or rendering partial settings. The self-hosted Pools heading no longer shows a misleading count after a failed load, and build pages keep their heading while loading or when a build is missing.
  • Fixes to pull requests and Slack. Opening a pull request before the agent pushes its branch now says to push the branch first, Bitbucket Data Center users with read-only access get a clear message asking for write access, and closing a pull request whose branch was deleted now explains that reopening restores it. Pull request status now updates when the pull request is merged or closed, and messages sent from a pull request page's sidebar name that pull request. In Slack, the agent no longer posts a contradictory access-denied reply in threads it already follows, and notices about skipped files list each file only once per thread.
  • Admin Fixes to agent visibility and conversation export. Owners can now switch an agent between private and team visibility even after the team's default visibility setting changes. Teams that export agent conversation content now also receive review subagent conversations.
  • Admin Routing rules have their own dashboard section. The routing rules that help the Slack bot pick a repository or environment for a request now live in a Routing section of the dashboard, with Team and Personal tabs, and old links redirect there.
  • Fixes to agent desktop security and network settings. A web page opened in the agent's own browser can no longer connect to the agent's remote desktop, and viewing the desktop from outside works as before. Agents whose environment has no network access mode set, or whose environment was deleted, no longer fail to start because their network access settings couldn't be resolved.

Integrations

GitLab

  • Admin Advanced setup for GitLab Self-Hosted instances. Admins can now open an Advanced section for each registered GitLab Self-Hosted instance on the Integrations page in the dashboard. From there, admins can save or reset a service account token, copy the webhook URL and secret for manual webhook setup, rotate the webhook secret, and see which webhook triggers are required. This is now available to everyone.
  • Fixes to GitLab.com token recovery. Starting an agent on many GitLab.com repositories no longer fails because one token expired or was revoked. Cursor now refreshes or re-creates the token and retries, even when GitLab.com reports the problem as a rate limit.

Microsoft Teams

  • Clearer messages for older Microsoft Teams link cards. In Microsoft Teams, only the most recent Link Account card can link your Cursor account. Opening an older card now shows that a newer card replaced it, or that your account is already linked, instead of a "linking expired" error.

Slack

  • A Slack workspace with a removed app shows Needs reconnect. In the dashboard, a linked Slack workspace where the Cursor Slack app was removed or its access was revoked now shows Needs reconnect instead of Connected, so you know to reinstall the app.
  • Fixes to Slack account linking. Linking your Slack account to Cursor no longer fails as already linked when the previous link belongs to a Cursor account on a blocked team. The new account takes over the link instead.

GitHub

  • Team repo pickers no longer hide your personal GitHub repositories. Team repo pickers now list repositories from your personal GitHub installation, even if you connected it under a different Cursor team.

Linear

  • Reconnecting Linear takes one step. On the Linear integration card in the dashboard, Reconnect now opens Linear's authorization window right away. The card refreshes when the window closes, even if you don't finish authorizing. If a Linear team can't be found, an All teams button takes you back to the team list.

Dashboard and admin

Security and audit logs

  • Admin Block personal API key creation for your team. Team settings have a new Allow personal API key creation switch under Security & automation. When it's off, nobody on the team, admins included, can create or rotate personal Cursor API keys, including CLI keys, automation webhook keys, and personal service tokens. Existing keys keep working until revoked. See members' API keys.
  • Admin Search your API keys and find fine-grained and legacy keys more easily. The API keys page now has a search box that finds keys by name or masked key, and on the Members tab, by member. Fine-grained and legacy keys appear in separate sections, each with its own sorting and pages. When you can't create a key, Create API key is disabled with a tooltip that explains why. See members' API keys.
  • Admin Track API key creation, revocation, and session exchanges in the audit log. API key create and revoke entries now show the key's name and ID, plus its service account when it has one. Revocations from rotating a key, deleting an automation or service account, disconnecting Jira, deleting an account, or revoking a leaked key now appear with the reason, and a new Session Exchange event records each time an API key or service token is traded for a session. See audit logs.
  • Admin See My Machines activity in the audit log. The audit log now records when a member's personal machine connects to or disconnects from Cloud Agents, and when a Cloud Agent starts on one, with the machine name, the agent, where it was launched from, and the disconnect reason. Filter for these with the new My Machines event type in audit logs.
  • Admin Trust the IP addresses your installed apps declare. Admins can now let installed apps connect from the IP ranges those apps declare, without adding each range to the team's inbound IP allowlist. The inbound IP allowlist settings list each installed app with its IP ranges, and each app's page shows the egress IP ranges it declares.

Plugins and MCP

  • Filter your personal plugins and MCP servers. In a team workspace, the Personal view of the Plugins page has a filter menu for type, category, and status (such as Connected or Disabled by admin). Filters stay in the page URL, so you can share or bookmark a filtered view.
  • Plugins removed from the marketplace are marked as removed. On the Plugins page, an installed plugin the marketplace no longer offers now reads "No longer in the marketplace", with Edit disabled and Disconnect still available. Removed plugins no longer appear in Suggested.
  • Admin See your team's first-party plugin policy at a glance. In workspace settings, the First-Party Plugins row shows whether all first-party plugins are enabled or only an allowlist, with a summary of the allowed plugins. Change opens a dialog to pick the mode and edit the allowlist.

Cloud Agents

  • Admin Track vulnerability findings from the Security Agents page. A new Remediation agents section on the Security Agents page groups vulnerability findings by stage. Open a finding to see its verification and patch timeline or to request verification or a patch, and admins can click Enable all to turn on remediation for every Vulnerability Scanner. See vulnerability remediation.
  • See how long each environment build took. The builds table on a Cloud Agent environment's page now has a Duration column that shows how long each finished build took. See builds.
  • Scope secrets to Origin repositories. When you limit a Cloud Agents secret to specific repositories, the picker now lists your Origin repositories alongside GitHub and other providers, with provider icons to tell same-named repositories apart.

Team settings

  • Admin Turn off voice chat for your team. Team settings have a new Voice chat section with an Enable voice chat switch that hides voice chat for everyone on the team when it's off. Voice settings also let you turn each status chime on or off and preview its sound. See turning voice chat off.
  • Admin Team privacy mode changes apply right away. When an admin changes the team's privacy mode, every member's next request now uses the new mode, instead of the old one for up to several minutes. The same applies when someone joins or leaves a team.
  • Admin Find any team setting from search or the command palette. Search on the Team settings page tolerates typos and matches any text in a section, not only its title. Admins can also type a setting's name in the Cmd+K command palette to open Team settings scrolled to that row.
  • Admin Members can see their team's restricted models. Team members now see the Restricted and early-access models section in Team settings, read-only, with whether the team acknowledged each model's data policy. Only admins can change acknowledgements.
  • Admin Clearer switching between repository blocklist and allowlist. The Blocklist / allowlist card in workspace settings shows which indexing mode is active. Change switches modes in a dialog that warns you before the switch removes the repositories already on your list.

Integrations

  • Admin Manage self-hosted GitLab webhooks and service account tokens. On the Integrations page, Manage instances for GitLab Self-Hosted now has an Advanced view for each instance where admins set or reset its service account token, copy the webhook URL and secret, and rotate the webhook secret. Rotations appear in the audit log. This is now available to everyone. See GitLab.
  • Admin Rollouts records Vercel deploys under their Vercel project. With Use my GitHub deployments on in Rollouts, Production and Staging deployments that Vercel's GitHub integration reports are recorded with the Vercel project as the service, and Vercel preview deploys are ignored. See sending deployment events.

Usage, members, and groups

  • Admin Filter team usage by service account, attribution, and client type. On the Usage page, admins of teams without unified billing can narrow usage to a service account, an attribution such as Cloud Agents, Automations, or Bugbot, or a client type. The chart, usage events, and CSV export follow the filters.
  • Admin Manage SCIM from a group's page. The actions menu on a group's page now includes Manage SCIM, which opens the same SCIM mapping dialog as the Groups list, or says setup is required when SCIM isn't set up. See group management.

Account and dashboard

  • Choose your home team. If you belong to two or more paid teams that aren't linked to an organization, Settings > Profile now has a Home team setting. It shows the team Cursor uses when you sign in and lets you change it or set it to None.
  • Clearer warning before deleting a merged account. If your Cursor account is merged with another account, the delete account confirmation now says that the merged account and its plan are deleted too, immediately and permanently.
  • Shared canvases open on their own page. Opening a published canvas from Shared resources now shows it on its own page with Copy link, Download as PDF, and Open in full screen. Older share links forward to the new page.
  • Switch themes from the command palette. Type "theme", "dark", or another mode name in the Cmd+K command palette to apply Light, Dark, or System, or pick Toggle theme. The palette also hides commands that would do nothing, like your current theme.

Fixes

  • Admin Fixes to MCP allowlists and plugins. A member's group MCP policy can no longer allow MCP servers or tools that the team's Cloud Agent MCP allowlist blocks, and plugins blocked by a team admin show "Disabled by admin".
  • Admin Fixes to members and spending. Organizations using org-level SCIM no longer have members' directory emails overwritten, which could remove them from mapped teams. Reset dates on the Spending page use your local time zone.
  • Admin Fixes to environments, integrations, and team settings. Secrets now appear on the configure page for environments configured from a file in the repository, and the Bedrock IAM role instructions name the correct permissions. Team invite links that carry tracking parameters now redeem, and a failed redemption says why.
  • Admin Fixes to dashboard navigation, old links, and the API keys page. Sidebar items open in a new tab with Cmd-click, Ctrl-click, or middle-click, and old links to Security Agents, Audit log, Data exports, and Deployed Agents open the new dashboard. When a key list fails to load, the API keys page says so and offers Try again when a retry can help, and copy buttons report a failed copy.

Automations

Security agents

  • Vulnerability Scanner can open fix PRs without a verification step. With Open fix PRs for confirmed findings on and Verify new findings off, new Vulnerability Scanner findings now go straight to a patch agent. The agent first writes a test that fails against the unchanged code, and opens no PR if it can't show the finding is real.
  • PR Approver waits longer for automated reviews. The PR Approver automation now waits up to 11 minutes, instead of 8, for configured Bugbot and Security Reviewer checks to finish before it approves or assigns reviewers. Fewer pull requests miss approval because a review check ran long.

Automations page

  • Sign in to an automation's MCP servers from its overview. The MCPs card on an automation's overview now shows the logos of its MCP servers and how many are connected. Click it to see which servers still need sign-in and authenticate each one. The status updates when you come back from the browser.

Tools

  • Images show inline in Origin pull requests opened by automations. When an automation opens a pull request on an Origin repository, screenshots and other images from the agent now appear inline in the description instead of as links to a sign-in page. Teammates who can read the repository and the agent see them without a public share.

Fixes

  • Automations on Origin repositories no longer fail when no branch is set. An automation on an Origin repository with no branch set now finds the repository's default branch and launches, instead of failing every run. This includes scheduled automations and automations owned by service accounts.
  • Fixes to the automation editor's pickers. Choosing an any-repo self-hosted pool no longer clears the repository you already selected. The repository picker in the desktop app now stays responsive with very large repository lists, as it already did on the web.
  • Fixes to vulnerability verification. The Vulnerability Scanner no longer starts a new verification session every minute while one is already running, and a reverify or patch request still starts one within a minute. Findings queued by a scan are now verified for up to 72 hours after they are queued, up from 24.

APIs

  • Admin Look up organization groups by slug, and add members by numeric user ID. In the Admin API, Organization Groups now include a slug, which is null when a group has none, and GET /organizations/groups accepts a slug query parameter that returns the group with that exact slug. When you add or remove group members, userIds also accepts the numeric id that GET /organizations/members returns, as a number or a string, alongside user_ IDs.

Security and approval agents

Security agents

  • Security reviews avoid quoting secret values. When a security review finds an exposed credential, its findings, inline PR comments, and Slack summary refer to the secret by its variable or setting name and point to its file and line instead of repeating the value, including partial, masked, or encoded forms.

Models and Cursor Router

  • Interrupted conversations no longer get stuck. If a response is cut off partway, you can now continue the conversation instead of getting an error. This affected models such as Gemini.

Cursor CLI

Auto-review

  • Hooks that ask for approval now prompt in Auto-review. When a beforeShellExecution hook asks for approval, the CLI now prompts you even if Auto-review would have allowed the shell command. A skip you chose yourself still skips the prompt.

Headless mode

  • Reasoning token counts in headless usage. Headless runs now include reasoningTokens in the per-turn and total usage of the result output when the model used reasoning tokens, so scripts can track the full token cost of a run.

Self-hosted workers

  • Admin Pool workers keep their API key away from agent tools. When agent worker runs as part of a worker pool, or the Cursor desktop app starts it, the worker removes its key from CURSOR_API_KEY before starting anything. Shell commands and tools that agents run can no longer read the worker's credential, and the key is redacted from tool output.
  • Self-hosted workers keep Namespace devboxes awake. When agent worker runs on a Namespace devbox, the devbox no longer stops as idle while the worker is serving an agent. The worker holds a Namespace task marker for as long as it runs and removes it when it exits, even after a crash, so an idle devbox can still shut down.

Fixes

  • Fixes to emoji and wide-character handling. Arrow keys, Backspace, Delete, and vim motions in the prompt editor now treat an emoji, flag, or accented character as one character instead of leaving broken fragments. Wrapped and truncated text, including the queued message banner, now lines up correctly with CJK characters and emoji.
  • Fixes to agent tools. Agent edits no longer hang on malformed patches, and searches with a match-all glob like **/* now run instead of failing. The agent reads workspace paths that contain repeated spaces correctly, tools the agent loads on demand show under their own name instead of as an empty MCP call, and long sessions no longer leak memory as the agent parses shell commands.
  • Better recovery from invalid tool arguments. When the agent calls a built-in tool it loads on demand with invalid arguments, the error now includes that tool's input schema, or for a very large schema tells the agent where to read it, so the agent can correct its arguments on the next try.
  • Image generation reports missing reference images. Image generation now stops with an error naming a missing or unreadable reference image instead of generating an image that ignores it.
  • Fixes to installing and launching the CLI. On Windows, when an update can't replace the launcher because an agent is still running, the background helper now stops after a set time and finishes on the next update instead of lingering. On macOS and Linux, the cursor command no longer finds itself on your PATH through a symlink or a trailing-slash directory, so it opens the Cursor app.

Pull requests

  • See what a pull request still needs from reviewers in origin pr view --checks. For an open or draft pull request, origin pr view --checks now prints a Review requirements section with the required gates, the approvals still needed, and the Code Review Gate, so you don't need to run origin pr requirements separately. Scripts can read the same data with origin pr view --json reviewRequirements, which is null for merged or closed pull requests. See view, diff, and checks.
  • See which reviews count toward a pull request. origin pr requirements has a Reviews on record section listing each reviewer's standing approval or change request and the pull request version it was filed on, so you can tell when an approval was given on an older version. Dismissed reviews are listed with when and why they were dismissed, and --json reviewsOnRecord returns the same data.

Stacks

  • origin stack push pushes the whole stack, and origin stack restack-remote asks before rewriting other people's pull requests. origin stack push now pushes every branch in the current stack that needs it, root to leaf, so you no longer need --all. Pass --only to push just the current branch. See push, move, and delete. origin stack restack-remote lists any pull requests in the range that you didn't write and asks before rewriting them; pass --yes to skip the prompt. Without a terminal or with --json, it refuses instead of waiting.
  • Stack commands on trunk explain what to do instead of failing. Running a stack command on your default branch now says that trunk isn't part of a stack and exits cleanly. origin stack get with no target fast-forwards trunk, origin stack up and top move into a stack based on trunk, and origin stack view, restack, and push exit without changing anything. See navigate.

Repositories

  • origin repo doctor sets up a checkout so plain gh works. The new origin repo doctor command points your checkout's origin remote at Origin and, for repos that GitHub owns, adds a github remote that gh uses as its default repository, so gh commands work without -R. git fetch and git pull stay on origin. origin repo clone sets up the same remote automatically, and origin repo set-origin-remote still works as an alias for origin repo doctor. See repositories.
  • Choose visibility when you create a repository. origin repo create takes --visibility internal or --visibility private, so you can make a repository private from the start. The default stays internal, and the output shows the visibility it set. See repositories.
  • Clearer origin repo clone-fast output and Windows clone fixes. origin repo clone-fast prints git-style Cloning into and Cloned into lines and plain-language messages when it falls back to git clone. Set CO_DEBUG to see timings and download details. If the checkout step fails, it keeps the downloaded repository and prints the git command that finishes the checkout. origin repo clone and clone-fast now set core.longpaths on Windows, and origin no longer exits silently without running. See CloneKit in CI.

Pull requests

  • PR file JSON shows where renamed files came from. origin pr view --json files and origin pr list --json files now include an oldPath for each file: the previous path of a renamed or copied file, or null. See view, diff, and checks.
  • Fixed origin pr list --search errors. When the server rejects a search, the error's carets point at the text you typed rather than a rewritten query. A query that names a state, such as state:all or is:merged, is no longer also sent with is:open when the CLI can't parse it. See list and status.
  • Clearer gate definition errors. When a pull request's gate definitions are invalid, origin pr requirements says so and names the policy file that defines the broken gates when that file is known. When gate definitions can't be read, origin pr gate says they're unavailable and gives the reason. See requirements and watch.

Cursor SDK

  • Run local agents on a file system you provide. Pass local.fs to a local agent, either a FileSystem or a factory called once per run, and its file tools (Read, Write, StrReplace, LS, Grep, Glob, Delete) and Task subagents work on that tree instead of your disk. Shell, hooks, and other tools that touch the host are turned off, and each run gets its own in-memory /tmp. Build trees with helpers like tempFs, unionFs, overlayFs, readOnlyFs, and childOverlayFs. TypeScript only.
  • Run the built-in file tools on your own file system. The new createFileSystemTools(fs, { workspace }) export gives you the native Read, Write, StrReplace, LS, Grep, Glob, and Delete tools on any FileSystem, so you can call them outside an agent run, for example from a direct tool-call endpoint, a test, or a CLI. Each execute call returns the same tool-call shape a run's tool_call event carries, and each tool comes with MCP-style annotations such as readOnlyHint and destructiveHint.
  • Compact a local agent's conversation with agent.summarize(). Local agents can now summarize their conversation on demand to free up context, like /summarize in the CLI. TypeScript only. The run streams summary messages as compaction progresses, and wait() returns the summary text. Calling it on a cloud agent throws ConfigurationError.
  • Give a local agent's own session custom executors and a workspace path. Set local.resourceProviderOptions and local.reportedWorkspacePath on Agent.create() to give a local agent's own session custom read, write, and shell executors and the workspace path its model is told to use, not only its subagents. Override either for a single send. TypeScript local agents only; setting them on a cloud agent throws a ConfigurationError.
  • Custom executors can now handle the Delete tool. Set overrideDeleteExecutor in local.resourceProviderOptions to run the agent's Delete tool through your own executor in place of the built-in delete. Your executor replaces the host's permission check and delete approval, so it decides what can be deleted. TypeScript local agents only.
  • Fixes to run streaming, error reporting, packaging, and Windows paths. Streaming a cloud agent run now reconnects and replays events when the stream is briefly unavailable, and results keep their duration and git info. Failed local runs report RunResult.error.code again, and the published package's ./agent, ./sqlite, and ./bundled entry points resolve to built files. On Windows, local runs no longer fail to start when the project sits at a long path.

Pull requests

  • origin pr edit no longer wipes a description by accident. It refuses an empty, whitespace-only, or null --body or --body-file; use the new --clear-body flag to clear a description on purpose. origin pr view and origin pr list return each --json field under the name you asked for, so --json body prints body, not description. Update scripts that read .description after asking for body.
  • Merge-when-ready is now called auto-merge. origin pr view, status, and requirements print an Auto-merge line, origin pr merge --auto and --disable-auto say auto-merge in their messages, and origin pr watch labels events [auto-merge]. Use --json autoMerge in scripts; --json mergeWhenReady still works but is deprecated. See merge.

Repositories

  • Faster CloneKit downloads on more systems. origin repo clone-fast uses its faster download engine with curl 7.75 or newer and GNU coreutils dd. When it can't, it prints a warning that says why and what to install, and the fallback engine uses bounded memory. See CloneKit in CI.

Web

  • Stacked pull requests move onto the base branch when their parent merges. In squash-only repositories, merging a stacked pull request now moves its open child pull requests onto the branch it landed on, instead of leaving them on the merged branch.
  • Commit co-authors show their name and avatar. When a commit's Co-authored-by trailer uses the email of the pull request author or another contributor to the repository, the commit row now shows that person's name and avatar instead of the raw trailer text. Co-authors who would appear twice are shown once.
  • Admin The audit log records signing key changes and ruleset bypasses on push. When a member adds or deletes a commit signing key, every team they belong to now gets an audit log entry with the key's name, type (SSH or GPG), and fingerprint. Repository push entries now list each ruleset rule a bypass let through, with the ref, ruleset, and rule type, and ruleset change entries identify groups that can bypass a ruleset by their public group ID.
  • Admin Rulesets match the same branches at merge time as at push time. Branch patterns in a ruleset's targeting, under Settings > Rules and Protections, now use Git-style pattern matching when Origin checks pull request merges, the same matching it already used to enforce pushes. A ruleset no longer covers a slightly different set of branches when someone merges than when someone pushes.
  • Admin Let installed apps through your IP allowlist from the addresses they declare. App managers can now list the IP ranges their app's servers call Origin from, under Egress IP Addresses when they edit the app. Team admins can then turn on Trust IP addresses declared by installed apps in the inbound IP allowlist settings, so each installed app's installation tokens are admitted from its declared ranges even when your allowlist doesn't cover them.
  • Clearer message when cloud agents need GitHub connected. If your GitHub account isn't connected in Cursor, the agent panel on a pull request page now says so and links to your Cursor integrations settings, instead of showing a generic error.
  • The pull requests list no longer shows "Approved" while a review requirement still blocks merging. On a repository's pull requests list, an open pull request now shows the Approved label only when no review requirement, such as a required code owner or approval count, still holds it, so the list matches the pull request page. When Origin can't evaluate the requirements, the list shows the review decision as before.

API

  • Agents can re-run flaky checks. The new rerun_checks MCP tool asks a check's CI app to run it again, either for one check run or for up to 20 failing re-runnable checks on a pull request, commit, or suite, and reports which runs it re-requested and which it skipped. checks_read now shows whether each run can be re-run, and pull request, review, and comment tools return people's display names and handles alongside their IDs.
  • Create private repositories from the Origin API. When you create a repository through the Origin API, you can now set its visibility to private so only people and groups you grant access can see it. Repositories are still internal by default, and public visibility is set later by updating the repository. See the Origin API changelog for details.

Fixes

  • Admin Fixes to code owner and review requirements. Approvals from people without review access no longer satisfy owner rules, and approving a pull request from Slack is refused if someone pushes to or closes it while the approval is being recorded.
  • Fixes to merge when ready. Turning on merge when ready as a service account, or through an app acting as a user, now fails with a clear error, and requests that could never merge are turned off with a timeline note explaining why instead of staying stuck.
  • Fixes to pull request pages. Opening a pull request's plain URL now always lands on the Activity tab, and the Activity timeline holds space for commits while they load, so it jumps less when they arrive. Code owner rows and team hovercards load faster. After you change a repository's visibility, the team access list on its Permissions tab updates right away.
  • Fixes to git pushes, clones, and LFS. Pushing a new branch to a repository with a linear-history rule no longer slows down in repositories with many refs, and clones right after a push that only moves branches or tags are fast again. Git LFS uploads that GitHub rate limits now return a retry-later response, so git lfs backs off instead of failing.
  • Fixes to GitHub mirroring. Branches Origin deletes after a merge are now removed from the GitHub mirror right away instead of lingering until the next push. A mirror push over HTTPS that hits a branch or tag deleted on Origin mid-push now runs again right away instead of failing.
  • Fixes to Slack notifications. Long pull request and comment bodies in Slack notifications are now shortened without breaking code blocks or opening collapsed sections.
  • Fixes to API errors and webhooks. Rate-limit and concurrency errors for API keys now point you to pull_request.* webhooks instead of polling. Fetching a pull request right after its webhook fires no longer returns not found, and webhooks no longer deliver events for a deleted repository.
  • Fixes to Bugbot on Origin pull requests. Retargeting or restacking a pull request now triggers a fresh automatic Bugbot review against the new base, even with no new commits.

Agents

  • Cloud Agent questions and sign-in cards no longer hold up the run. When a Cloud Agent asks you a question, needs you to sign in to an MCP server, or asks for access, it now ends its turn and leaves the card open until you answer, even after a reload. Sending a message instead skips a pending MCP sign-in card so the agent keeps going, and an agent working toward a goal waits for your answer instead of continuing on its own. This is now available to everyone.
  • Slash commands in Cloud Agent prompts run the command you meant. A slash command followed by punctuation, like /review, or /deploy?, now runs as the command instead of being ignored. A command or skill you pick from the menu is kept instead of being swapped for a team or repo one with the same name.
  • Admin Clearer errors for spend limits and AWS roles. When a spend limit blocks a Cloud Agent, the error now says whether your limit or your team's was hit, shows its dollar amount, and tells you to raise it in team settings or ask an admin. If the AWS IAM role in CURSOR_AWS_ASSUME_IAM_ROLE_ARN can't be assumed, the run shows an "AWS IAM role could not be assumed" warning with the reason, and AWS commands in the run print that reason instead of a missing-file error.

MCP

  • Admin The Cloud Agent Allowlist now covers every MCP server a Cloud Agent uses. When admins turn on a separate Cloud Agent Allowlist, it now also governs MCP servers loaded from the agent owner's settings, inline MCP configs passed when starting or following up on an agent, and built-in connectors, including its disable-all setting. An environment MCP rule set to read-only now allows no tools from that server instead of all of them.
  • Admin Clearer messages when your organization's Microsoft sign-in policy blocks an MCP connection. When you connect an MCP server that signs in with Microsoft and your organization's policy refuses the connection, the sign-in page on cursor.com now leads with what to do, such as asking your admin to allow it or signing in again and finishing every step Microsoft asks for. This is now available to everyone.

Pull requests

  • Agent commits reliably credit you as co-author. Commits made by Cloud Agents now always carry your Co-authored-by line, even when the agent commits with --no-verify or the message already names another co-author. Cherry-picks and rebases are left untouched, and repositories checked out as worktrees or submodules now get the same co-author line.
  • Multi-repo pull requests target your starting branch. When a multi-repo agent opens a pull request on the repository the run started from, it now targets the branch you started on instead of defaulting to main. Pull requests on the other repositories in the run still use the base branch the agent picks.
  • Cloud Agents push once they finish testing instead of after every commit. Agents still commit as they work, but now push the branch when testing is done rather than before testing and again after, so fewer CI runs start on your repository. The agent opens or updates its pull request after that push.

Environments

  • Cancel in-progress environment builds, from an agent or as any team member. A Cloud Agent can now cancel an in-progress build of its own environment when you ask it to, freeing a build slot. On teams where only admins can change environments, any member can now cancel a draft build of a team environment from the Builds menu on the Cloud Agents dashboard. See Manage Builds.

Integrations

  • Turn off Slack link previews for links you post. A new Slack link previews switch under Pull requests in your account settings stops Cursor from unfurling the pull request, code, and artifact links you post in Slack. It's on by default. Previews also stay quieter: a link previewed in a channel in the last 30 minutes doesn't unfurl again, and messages with three or more Cursor links get no previews. See Slack link previews.

Web

  • Subagent pages on the web point you back to the parent agent. When you open a subagent that a Cloud Agent started, the follow-up box is replaced with a note that subagents can't take follow-ups and an Open parent agent button that returns you to the task in the parent's conversation. Opened subagents also no longer show up as separate agents in the sidebar.

Fixes

  • Fixes to environment builds and repository cloning. Cloud Agents now reuse an environment build's existing checkouts when the requested repositories differ, clone only what's missing, and no longer fail with a repo config mismatch on pinned environments. Forks keep the source agent's workspace root, and follow-ups of no-repo agents whose blank environment was deleted start on a replacement. Clones that time out or stall are retried, and deployments with private CA certificates can check out repositories without TLS errors.
  • Fixes to pull request handling. Agents can now comment on, reply to, and resolve review threads on GitHub pull requests in setups where this used to fail, keep a pull request ready for review when your "Open PRs Ready for Review by Default" setting is on, and find the right branch when opening a pull request from a single-repo workspace. Agents no longer miss the final CI result after re-subscribing, and Fix with agent on the pull request page now names the base branch or failing checks and shows a working indicator.
  • Fixes to Projects. Moving a Project's notes to tracked tasks no longer skips plain checkbox checklists or tasks above the first status heading, and handles up to 3,000 tasks.
  • Fixes to MCP servers. MCP sign-in cards, permission errors, and rate-limit or policy messages now name servers by their readable name instead of an internal ID, and a tool refused for an ungranted permission no longer disconnects the server. Finishing an MCP sign-in on the web refreshes status in every open tab and returns you to the page you started from. Linear connections are no longer marked as needing reconnection when reconnecting could not fix the error.
  • Fixes to Self-Hosted Machines. When your machine goes offline, even mid-turn, the agent now says it may be asleep and tells you to wake it or run agent worker start, and recurring wakes no longer start turns that fail right away. Slow Self-Hosted Machines no longer end a turn with an unreachable-environment error, and the remote desktop of Windows Cloud Agents now opens on dedicated deployments.
  • Fixes to the Cloud Agents web dashboard. Agent conversations load and prompt attachments upload in older browsers (Chrome and Edge before 116, Safari before 17.4), and environment build logs no longer show raw color codes. Saving an environment setup whose environment was moved or deleted now explains what happened and what to do next. Self-hosted GitLab and Bitbucket repos show their real provider in the repository picker, and missing GitHub connections now show a prompt to connect instead of a generic error.
  • Fixes to agent runs. Agents that run out of disk space now stop instead of retrying, with a message suggesting you ask the agent to delete large or unneeded files, and edits from a failed or retried model response are no longer applied. Follow-ups sent while an agent is waiting are no longer dropped, and artifact files still open after the agent's machine is gone.
  • Fixes to Slack and Microsoft Teams integrations. Cloud Agents in Microsoft Teams now work in older channels, and the public Install Slack app link works again for signed-out visitors. The Slack welcome message no longer repeats on every launch in Enterprise Grid workspaces.

Integrations

GitLab

  • Deploy agents from GitLab Self-Hosted. You can now deploy agents from repositories on a GitLab Self-Hosted instance connected to your team, including instances served under a URL path such as corp.example.com/gitlab. New Deployment lists those repositories, and each deploy checks clone access first. If Cursor can't reach the GitLab host or the project token can't read the repository, the deploy stops with an error that names the problem.
  • Admin Clearer GitLab access errors on the Integrations page. When Cursor can't reach a connected GitLab or GitLab Self-Hosted integration, its row on the Integrations page now explains the problem and how to fix it, with a Try again or Reconnect button, instead of showing a plain connected or disconnected state. On a team's Integrations page, if your own GitLab account is linked but the team's isn't, the GitLab row says so and names the account you're connected as.

GitHub

  • Admin Connect GitHub Enterprise for yourself. On a team's Integrations page, a GitHub Enterprise host now offers Just for me next to Connect, as GitHub already does. Members who can't manage the team's integrations see only Just for me, with a note that a team admin can connect it for the team. Disconnecting GitHub or GitHub Enterprise on a team now explains that only your sign-in is removed and the team's app stays installed, and a signed-out row shows a Sign in button.

Bitbucket

  • Admin Fixes to Bitbucket Data Center. Repository lists on large Bitbucket Data Center instances no longer repeat repositories or miss later ones. Adding a user when the server has no licenses left now reports a seat limit instead of an unexpected error.

Integrations page

  • Open self-hosted integrations by clicking anywhere on the card. On the Integrations page in the dashboard, if you can manage integrations, clicking anywhere on the GitHub Enterprise, GitLab Self-Hosted, or Bitbucket Data Center card now opens its management page. Buttons on the card keep their own actions.

Dashboard and admin

Cloud Agents

  • Admin Manage deployed agents from Cloud Agents. Once your team has deployed an agent, Cloud Agents > Deployed agents lists your team's deployed agents with their status, channels, 7-day usage, source, and last deploy, and New deployment deploys an agent project from a repository at any branch, tag, or commit. Each agent's page shows sessions with step-by-step traces, logs, and deploy history, and lets you roll back, redeploy, stop, or delete the agent. Deployed agents is in beta.
  • Admin Members can cancel draft environment builds when environment edits are admin-only. On teams that limit Cloud Agent environment edits to admins, any member can now cancel an in-progress draft build of a team or repository environment from the builds table or a build's page. Canceling other team builds and activating drafts still need an admin. See managing builds.

Plugins and MCP

  • Admin Browse and install plugins from the Plugins page. Add on the Plugins page now opens a menu with Plugin and Custom MCP. Plugin opens a searchable catalog with Featured and category sections that installs in place, and Custom MCP opens the Add MCP server form, grouped into Where it's added, Connection, and Authentication. See the marketplace.
  • Admin Plugins your team blocks no longer offer Add. In the plugin marketplace on the Plugins page, plugins your team's policies would refuse to install, such as ones whose MCP servers an admin disabled or that aren't on the team's allowlist, now hide Add instead of failing after you click it. This is now available to everyone.

Security and audit logs

  • Admin Stop members from creating personal API keys. Team admins can turn off Allow Personal API Key Creation in the new Personal API Keys card on the API keys page. Members then can't create or rotate personal Cursor API keys, and existing keys keep working until revoked.
  • Admin More audit log events, and a searchable event filter. The audit log now records cloud automations being created, updated, enabled, disabled, or deleted, members setting or leaving a home team, organization API key revocations, Origin repository renames, and commit-signing keys being added or deleted. Login entries name the client and version that signed in, and the event type filter has a search box. See audit logs.

Members and groups

  • Admin Move admins who don't use Cursor to a Free seat. On teams with seat tiers, the seat menu on the Members page offers Free for admins who manage the team but don't use Cursor. A confirmation explains that Free admins can't use Cursor and that the change applies right away. See seat tiers.
  • Admin Review seat upgrade requests from the Members page. On teams with seat tiers, admins get an Upgrades tab on the Members page that lists members' pending seat upgrade requests and, on self-serve teams, recommended upgrades. Each row has Upgrade, and a bulk button upgrades all seats or the ones you select. The tab isn't offered during a Teams trial. On these teams, the Members page also shows when member usage resets, in your local time zone. See changing a seat tier.
  • Admin Legacy SCIM groups appear on the Groups tab, and synced groups show their last sync. The Groups tab lists legacy SCIM directory groups with a Legacy badge, including for teams without an organization. Each synced group's page shows when it last synced, and when you add an organization group you can choose Manual or Synced and pick a directory group to sync from. See group management.
  • Admin Clearer spend limit controls. Lowering a member's spend limit warns you when a higher group or team limit still applies, and Cloud Agents blocked by a limit say which team or member limit stopped them.

Team settings and integrations

  • Admin Team members can open Team settings. Every team member now sees Team settings in the dashboard navigation. Members who can't manage settings get a read-only view of the team's name, privacy mode, and Origin review and repo creation providers.
  • Admin Connect GitHub Enterprise just for yourself. On a team's Integrations page, GitHub Enterprise now offers Just for me next to Connect, and members who can't manage integrations see that a team admin can connect it for the team. Disconnect on GitHub and GitHub Enterprise now explains that it removes only your own sign-in while the team's app stays installed. See GitHub.
  • Turn off Slack link previews. A new Slack link previews switch in your pull request settings shows or hides previews of the pull request, code, and artifact links you post in Slack.
  • Filter Insights to weekdays. The Include weekends switch on the Insights page is now a Filter by day dropdown with All days and Weekdays only. Existing links keep their filter. See filtering analytics.

Fixes

  • Admin Fixes to the Members, Usage, and Groups pages. The Members page keeps listing members, with search and sorting, when usage data fails to load. Directory group spend limit overrides on pooled and Enterprise token plans now save to the monthly limit that is enforced, and group pages show clear empty, not-found, and load-failure states with Try again.
  • Admin Fixes and polish. Settings save instantly and offer Try again when a save fails on a temporary network or server error, and team rule globs like *.{ts,tsx} stay as one pattern. Adding or renaming an MCP server can no longer overwrite one with the same name, and long emails in the audit log are truncated, with the full address on hover.

Automations

Automations page

  • A refreshed Automations list with icons, run history, and templates. The Automations list now shows each automation's icon, a 15-day run sparkline, and a count of what you're viewing, with Sort (name, last updated) and Filter (type, creator, status) menus. You can give an automation its own icon and color in the editor, and start from a template in the Templates section with one click on Use. Now available to everyone.
  • Search for a model when editing an automation. The model picker in the automation prompt editor now has a search box, so you can type to filter the list instead of scrolling. Now available to everyone.

Security agents

  • Admin Vulnerability Scanner can verify findings and open fix PRs. Each Vulnerability Scanner automation now has its own Verify new findings and Open fix PRs for confirmed findings remediation settings, now available on Teams and Enterprise plans. Verification agents write tests to confirm each new finding, and patch agents open fix PRs for you to review. Both settings start on when an admin creates a scanner, and only team admins can change them.
  • Security agent pages get clearer names. Under From Cursor in Automations, the Security Reviewer page is now Security Review and the Vulnerability Scanner page is now Vulnerability Remediation, with new page addresses to match. Old links, including links to a specific agent or finding, redirect to the new pages. On the Vulnerability Remediation page, the sections are now Scanner Agents and Remediation Agents, and Remediation Agents sits above Recent Runs.
  • Vulnerability Scanner keeps secret values out of its reports. When the Vulnerability Scanner finds an exposed credential, its findings, Slack reports, and fix pull requests name the secret and point to the file and line where it lives, without repeating the value or any part of it. Fix pull requests that remove a hardcoded secret also remind you to rotate it, since removing it from the code doesn't revoke it.
  • Security Reviewer runs on GitLab merge requests again. Fixed an issue where Security Reviewer automations matched a GitLab merge request, including on self-managed GitLab, but never ran and left Recent Runs empty.

Automation runs

  • Run history shows skipped scheduled runs. When a scheduled automation fires while its previous run is still in progress, run history now records a skipped run that explains why, instead of showing nothing.
  • Clearer messages when a run can't start. When a run fails because your machine was offline, run history now shows Machine offline with the machine's name and the agent worker start command to bring it back. A run whose selected environment was deleted or is no longer accessible shows Environment not available and asks the owner to select an environment and save. Runs blocked by a Cursor-wide rate limit now say there is nothing to change on your side and to trigger the run again shortly.

Fixes

  • Fixes to the automation editor. Adding or renaming an MCP server from an automation's actions no longer overwrites an existing server with the same name, and the repository picker stays responsive with very large repository lists.
  • Fixes to automation runs. Automations that approve pull requests, resolve review threads, or manage check runs no longer fail with an "Unknown action type" error before the agent starts. This applies to CI, scheduled, webhook, and GitHub triggers.

APIs

  • Admin Verify and patch security findings from the API. Use the Vulnerability Remediation API with a team API key to list your team's Vulnerability Scanner findings with GET /security/findings. Start verification agents and fix pull requests with POST /security/verifications. Track them with GET /security/verifications and GET /security/verifications/{verificationId}, which return each verdict, its patch status, and links to the fix pull request and Cloud Agent. Service accounts get two new scopes for these endpoints: View security findings and Verify and patch security findings.

Security and approval agents

Security agents

  • Admin Verify vulnerability findings and open fix PRs automatically. Remediation Agents are now available on Teams and Enterprise plans. Admins can turn on Verify new findings to test each finding a Vulnerability Scanner reports with an exploit test, and Open fix PRs for confirmed findings to get a fix PR, ready for review, for each confirmed issue. Each scanner keeps its own settings, and new scanners created by admins start with both on. Enterprise teams can also request verification and fixes for specific findings through the Vulnerability Remediation API.
  • Admin Security dashboard pages renamed to Security Review and Vulnerability Remediation. Under From Cursor in the dashboard, the Security Reviewer and Vulnerability Scanner pages are now Security Review and Vulnerability Remediation, and old links redirect to them. The Remediation Agents section on Vulnerability Remediation is marked Beta.
  • Fixed duplicate findings and slow repository pickers on the security dashboard. The Remediation Agents section no longer lists the same scanner finding twice. Links to Vulnerability Remediation, such as those from Slack, now scroll to the Remediation Agents section once the dashboard loads. The repository picker on security and approval agent triggers stays responsive with thousands of repositories.
  • Fixed follow-up security reviews skipping fixed issues. When you push to a pull request again, Security Review marks issues you've fixed as fixed instead of skipping them without notice.

Approval agents

  • Fixed approval reviews failing when rules share a prompt. When several matched rules use the same prompt text, PR Routing & Approval now includes that text once, so the repeated text no longer makes the review too large to run.

Reviews and Autofix

  • Bugbot reviews Origin pull requests again after a retarget or restack. When an Origin pull request's base changes without a new commit, for example after you retarget it or restack it onto main, Bugbot now runs a fresh automatic review against the new base. Previously Bugbot skipped the run because it had already reviewed that commit, which left some pull requests with only a stale "too large to review" result.
  • Fixed Autofix skipping runs when your GitHub account is linked to several Cursor accounts. On a personal Bugbot install, Autofix now prefers your personal Cursor account and its privacy settings over a team account you also belong to. Previously Autofix sometimes checked the team's privacy settings instead of yours and skipped the run.

Fixes

  • Fixes to review reliability during GitHub errors. Bugbot reviews no longer fail when GitHub briefly fails or rate-limits requests for a new access token.
  • Fixes to the Bugbot dashboard. Hovering a clickable row at the top or bottom of a dashboard card no longer spills past the card's rounded corners.

Pull requests

  • origin pr create opens pull requests ready for review. New pull requests open ready for review by default instead of as drafts; pass --draft or --status draft for a draft. If your 'Never open PRs as drafts' setting or the repository's setting turns a requested draft into a ready pull request, origin pr create says why. See create.
  • Agentic code owner holds read like classic approval requirements. When a ruleset uses Agentic code owner review, origin pr requirements lists its holds under "Approvals still needed" in the same shape as classic requirements, such as "1 approving review from anyone but the author (0 so far)".

Stacks

  • origin stack sync cleans up merged branches it used to miss. It deletes the local branches of merged and closed pull requests even after origin stack get stopped tracking them, unless that would lose a commit. A branch renamed with git branch -m is kept and reported with the command that lets sync delete it. Stack commands run on a merged or closed branch point you to origin stack sync.
  • Stack commands no longer throw away a rebase you're resolving. origin stack move, delete, restack, and sync stop with an error when a rebase or merge they didn't start is in progress, instead of aborting it and discarding your resolved conflicts. When a stack rebase hits a conflict, the message names the branch, its new parent, the conflicting files, and the commit being applied, then lists the commands to run next.
  • Clearer origin stack messages. origin stack sync, get, push, restack, and delete-branch say exactly what changed, for example moved b's parent from main to a, as on Origin. origin pr create and origin pr edit explain when a branch's stack parent couldn't be saved in your clone and point you to origin stack get.

Scripting

  • Stack commands take --json field lists. origin stack commands such as get, restack, and push take a comma-separated field list with --json, like origin stack get --json members,checkedOut, and a bare --json is a usage error that lists the fields. See JSON and exit codes.

Fixes

  • Fixes to origin stack get with restacked branches. It fast-forwards branches that Origin restacked instead of reporting them as diverged, including in fresh clones, with core.logAllRefUpdates=false, and with log.showSignature enabled.

Cursor mobile

Voice calls

  • Voice calls show on the Lock Screen and keep the chat focused. When iOS can't show a voice call on the system call screen, Cursor for iOS shows it as a Live Activity on the Lock Screen and Dynamic Island, where you can mute or hang up. During a call, the chat hides your transcribed speech and Worked sections with nothing in them, so you see only the agent's replies, questions, plans, and edits. Call cards appear in the conversation where the call started, not pinned to the bottom.

Account and billing

  • Manage Plan when you hit included usage on the top plan. If you're on the highest personal plan and use up your included usage, the Included Usage Reached message now includes a Manage Plan button.

iPhone Duo

  • The sidebar and chat split at the fold on iPhone Duo. On iOS 27.1 or later, when iPhone Duo is half open in landscape, the agents sidebar ends at the hinge so the sidebar and chat each sit on their own half of the screen instead of straddling the fold. If either side would be too narrow, the usual layout stays.

Fixes

  • Layout and visual polish on iPhone and iPad. The composer and the iPad sidebar's account row sit on the bottom safe area and line up with each other, without an extra gap above the home indicator. Toolbar icons match the size of native toolbar symbols, the sheet grabber shows only on resizable sheets, and the Access Disabled screen keeps a readable width on iPad.
  • Conversations and the repo picker. The app no longer crashes on iOS 27 when loading conversations or the profile screen. A lone Thinking step in a running conversation updates to show how long it took once it finishes, and the repo picker no longer lists or preselects unpublished Start from scratch drafts.

Agents Window

Projects

  • Project split panes keep your focus and come back after a reload. In a Project's split panes, the pane you focused stays focused when panes return or the window reloads, and the editor panel follows whichever pane you focus, including chats dropped in from outside the Project. A chat you opened in another Project's split reopens in that split after a reload or Back/Forward, and the Project's own chat tab stays pinned first in its tab strip.
  • Calmer start for new Projects. While a new Project sets up, its conversation now shows a single "Setting up project" row until the agent's first output, instead of reflowing through environment setup steps. A new Project starts with its apps panel closed, and the Project app opens on its own once the first task appears.
  • Track a Project's tasks in List and Board views. Starting in 3.26, a Project's notes are a task tracker that the coordinator keeps up to date, with each task's status, ID, agents, and pull requests, shown as a List or a Board. The Project tab stays pinned first in the right panel, and its header switches between List, Board, and Library. While the Project's files sync, the tab shows a loading state, or an error if sync fails, instead of looking empty.
  • Filter tasks in Project notes. In a Project's Notes tab, a Filter button next to find narrows the task List or Board by Status, PR, Tags, or Owner. Active filters show as chips you can edit, switch between "is" and "is not", remove, or clear, and typing in the find bar filters tasks by text too.
  • Tasks with several PRs stay readable in Project notes. In a Project's task List, a task whose PR chips would crowd its title shows a single "N PRs" pill with an icon for each PR's state. Click the pill to open a menu of that task's pull requests.
  • Find repositories, folders, and environments faster in the project picker. The project picker opens on a search field, with your repositories, folders, and cloud environments in one list and owners shown on search results. Paste a clone URL to clone it from the picker, or use the footer to clone a repository, start from scratch, or create a new repository.
  • Project workers stay on their assigned task. When a Project worker runs a plan handed down from its coordinator, it works on its own task instead of acting like the Project's coordinator.

Chat and composer

  • Fixes to prompt attachments. Attached images and videos in sent messages are labeled with their file name, and images with very long file names attach instead of failing to save. Adding a HEIC or HEIF image on Windows or Linux now explains that those images need macOS, and slow HEIC conversions get more time and a clear timeout message. Pasted documents with no local file attach to cloud agent prompts, and videos or documents that can't be attached show a warning instead of disappearing.
  • More reliable queued messages and action buttons. Action buttons that send a message, such as git actions, now follow your queue or steer behavior and keep your typed draft and attached context. Send Now on a queued message you just edited sends your saved edit from any composer, and shows an error instead of sending stale text if the save fails. Delivered messages from other agents no longer linger in the queue.
  • New chats no longer pick up a Project's environment. Starting a new agent while a Project chat is selected no longer runs the new chat in that Project's environment. With the new-agent default set to Last Used, New Chat uses the environment of the last chat you used outside a Project, including when a Project worker was opened before its Project finished loading.
  • Your messages grow with your chat width. Your message bubbles in agent chats are no longer held to a fixed maximum width, so widening the Chat Width setting gives long messages more room. In a Project's coordinator chat, messages keep a fixed maximum width.

Navigation and layout

  • Tabs keep your place across agent switches and restarts. Tabs in the Agents Window now return to where you left them after you hide the panel, switch agents, or restart. Files reopen at the same PDF page and zoom, image zoom and pan, CSV scroll, and code scroll and cursor; the Changes tab keeps its scope, branch, and picked commits; canvases keep their Rendered or Source view; and expanded folders, open find bars, and agent-opened tabs like Agent Changes and Desktop come back too.
  • The sidebar remembers which sections you collapsed or expanded. Collapsing the Pinned section or the Repositories, Workspaces, or Chats header now sticks across reloads and restarts, and sections you expanded with More keep showing the same number of rows. A repository or workspace section you expand stays open after a reload, even if its agents are idle, until one of them has new activity. Starting a new chat also unfolds a collapsed Repositories or Workspaces header so the chat is visible.
  • Edit > Undo and Redo now act on what you were working in. Undo and Redo from the menu bar apply to the prompt, editor, terminal, or browser that had focus before you opened the menu. They undo window actions like archiving an agent only when nothing editable has focus, so typing in the prompt and then archiving from the sidebar no longer sends Undo to the prompt.

Files, changes, and pull requests

  • Clearer merge actions on pull request tabs. When a merge on a pull request tab lands a stack of several pull requests, the merge button shows how many it will land.
  • See how stacked pull requests build on each other in the PRs tray. The PRs tray in a Project groups stacked pull requests on a rail, so you can see how each PR builds on the one below it. Forked stacks branch off the parent line, and the tray's stack view matches the stack navigator in the PR tab.

Customization and settings

  • Hide the unread count next to the menu bar icon. On macOS, the menu bar menu has a new Show Count option. Turn it off to keep the menu bar icon without the unread agent number beside it. It's on by default and persists across restarts.

Browser

  • Preview pages on iPhone Duo and rotate any device in the browser. The browser's responsive design mode adds iPhone Duo (Closed) and iPhone Duo (Open) device presets. Rotate works for every preset, including Laptop and Desktop, and the rotation carries over when you switch devices.

Fixes

  • Fixes to sidebar keyboard navigation. Arrow keys, Home, and End now move through every reachable row and button in the sidebar, skip hidden groups and hover-only buttons, and no longer open a preview card on each row. Keyboard focus stays in place when you pin, unpin, move, archive, or collapse agents, or press More. Screen readers also stop announcing repository headers as keyboard-reorderable.
  • Fixes to loading agents and Project workers. Opening a finished cloud subagent from a link or by id no longer fails because it was unloaded first. Agents hidden from the sidebar now load their tiles instead of waiting indefinitely, and a Project no longer shows as running because of workers that already finished.
  • Fixes to transcript and prompt styling. Steer messages and Send now interrupts no longer appear grayed out while they wait for the agent to pick them up. Time markers in the transcript now show as a label between two lines, lined up with message bubbles.
  • Fixes to voice calls. Voice calls in the Agents Window no longer error out when the agent you're talking to is closed, unloaded, or hasn't loaded its conversation yet. Clicking Stop or Dictation twice, or double-clicking, no longer starts a voice chat by accident, and the requests a call sends to the agent no longer show up as chat messages in the transcript.
  • Fixes to agent reliability. On Windows, hooks now get clean UTF-8 payloads, and screen readers announce the Restore Checkpoint control in chat.
  • Fixes to Projects. Opening a newly created Project no longer briefly fails to load its tasks and files, opening a Project whose parent conversation was unloaded shows that conversation instead of an error screen, and Project tabs you return to restore their layout and scroll position. A Project split showing a chat from outside the Project no longer collapses when another view briefly takes the panel, and long task lists in hidden tabs no longer crash. The coordinator keeps its instructions on turns started by notifications, and unpublished Start from Scratch drafts no longer appear in the project picker's Recents.
  • Fixes to browser Annotate mode. Pressing Cmd/Ctrl+L while editing page text in Annotate mode adds the edit to the chat input instead of queueing it. Picked elements show as inline mentions in notes again.
  • Fixes to subagent rows. Subagent rows in an agent's conversation no longer turn into empty rows of another tool when a late update arrives for the same call.
  • Fixes to the Automations repo picker. The repository picker no longer loads indefinitely for large Bitbucket Cloud accounts. If loading takes more than 25 seconds, it shows the repositories found so far.
  • Fixes to Projects and setup banners. Opening a Project whose parent agent hasn't loaded yet now lists its known cloud agents instead of failing. Collapsed sections in a Project's task List no longer show an extra-wide gap, and the Dismiss tooltip on environment setup banners now shows that Esc closes them.

Cursor IDE

Editor and app

  • Reduce Motion now covers Cursor's own UI, and drop overlays snap into place. With workbench.reduceMotion on, Cursor's spinners, loading bars, menus, and dialogs pulse or fade instead of spinning or scaling, even when your OS preference is off. When you drag an editor or terminal to split it, the drop overlay snaps straight to each target instead of sliding.
  • Expand, zoom, and copy Mermaid diagrams in markdown preview. Mermaid diagrams in markdown preview now have the same controls as diagrams in chat. Open one in a larger view to zoom around it, or copy its source.
  • Cursor opens its first window faster. Cursor now waits until the first window is open before it loads error reporting, so that work no longer delays a cold start.

Agent and chat

  • Agent no longer garbles non-UTF-8 and non-ASCII text. Agent terminal output no longer garbles non-ASCII characters split across output chunks, and file reads and edits keep the original bytes when a file has some invalid characters. Codebase indexing now reads Latin-1, Windows code page, and legacy CJK files as text, so semantic search finds matches in them.
  • Hooks can tell which agent spawned a subagent. Hook scripts now receive parent_agent_id, the conversation ID of the agent that spawned the subagent, on subagentStart and subagentStop, so scripts no longer have to rebuild the agent tree from earlier events.
  • Agents keep working on remote machines that can't reach Cursor. In Remote-SSH, WSL, and container windows, agents now route requests through your local machine when the remote can't reach Cursor's servers. If the agent connection drops mid-turn, the chat shows Reconnecting... and resumes the turn instead of failing.

Fixes

  • Fixes to agent links in chat. Links to cloud agents in chat now show the agent's name as soon as its details load, instead of waiting on an unrelated update.
  • Fixes to MCP. When an MCP server asks for input, the form in the agent chat now scrolls so its Submit and Cancel buttons sit above the prompt box, and the buttons are reliably clickable. MCP sign-ins whose OAuth refresh token has expired now ask you to reconnect instead of failing over and over.
  • Clearer Add to Chat button in Review Changes. The Add to Chat button in the Review Changes file list now uses a chat bubble with a plus icon.
  • Fixes to Linux package update notifications. On Linux installs from the deb or rpm package, Cursor no longer shows update notifications when apt or dnf has nothing newer to install. Cursor now checks its package repository before notifying you, and shows one notification per new version, however many windows you open or restarts you do.
  • Fixes to non-ASCII text on Windows. Python-based MCP servers started with python, py, uvx, uv run, or pipx run no longer garble or fail on non-ASCII text.

Cursor CLI

Cloud Agent environments

  • Update an existing environment with agent env deploy --env. Pass --env <id>, using the public ID that agent env list shows, to redeploy an environment file onto an existing Cloud Agent environment. The environment keeps its name and repositories, so --repo becomes optional. If --name, --repo, or --scm-repo-node-id don't match the stored environment, the command stops with an error that explains the mismatch.

Self-hosted workers

  • Self-hosted workers run under the team you selected. agent worker start and agent worker debug now register and check the worker under the team you picked at agent login or with agent team switch, instead of your account's default team. If that team doesn't allow you to use Cloud Agents, the worker exits with a message naming the team and suggesting agent team switch.
  • Fixes to self-hosted workers. Workers sharing a machine no longer fail to start the editor server when another worker is running an older build of it, and workers answer editor connections right away while that server downloads instead of hanging. Connections a worker can't serve now fail immediately, backend connections no longer drop when the machine is busy, and the first turn of a session no longer misses the skills of the person who started it.

Subagents

  • Admin Custom subagents no longer fail when their model can't run. When a custom subagent is pinned to a model your team's admin settings block, that requires Max Mode, or isn't available, it now runs on the parent agent's model instead of erroring out. The subagent's result notes which model it ran on, so the agent knows the pinned model was skipped. See when the configured model won't be used.
  • Composer 2.5 subagents run and bill at the speed you picked. Subagents set to Composer 2.5 now run and are billed at standard speed instead of the fast tier. Subagents on Composer 2.5 Fast run at fast speed without failing with an "Unknown model ID" error.
  • Fixes to subagents. When the agent waits on a subagent that failed or was stopped, it now sees the failure instead of treating the subagent as finished.

Rules

  • Nested AGENTS.md rules now apply when the agent edits a file. When the agent edits or creates a file in a subdirectory, it now sees that directory's AGENTS.md and nested always-apply rules, even if it never read a file there first. Each rule is shown once per conversation, and glob-scoped and root rules behave as before.

Fixes

  • Fixes to background summarization. Long agent sessions now keep summaries computed in the background instead of discarding them and summarizing again.
  • Fixes to tool call display. Edit, shell, and task tool calls now show as started while the agent is still writing them.
  • Fixes to agent image generation and web fetch. Image generation prompts written in Chinese, Japanese, Thai, and similar scripts are no longer rejected as too short. Web fetch no longer fails on long pages when the cutoff lands inside an emoji, and video generation now explains why it couldn't read an input image.
  • Fixes to Auto-review. Auto-review of MCP tool calls now sees the tool's input schema when the server sends it as a JSON string.
  • Fixes to MCP servers. A single invalid tool or a tool schema without a type no longer hides the rest of an MCP server's tools, and a server that fails to list its tools now reports an error to the agent.
  • Fixes to Windows. Shell commands the agent runs in PowerShell now use UTF-8, so non-ASCII output no longer appears garbled. Staged updates are no longer held back by leftover markers from crashed or killed processes, and the background update check no longer flashes console windows.

Stacks

  • Clean up merged stacks with origin stack sync. It fast-forwards trunk, deletes your local branches whose pull requests were merged or closed, and restacks the branches that sat on them. Use --dry-run to preview or --no-restack to leave branches in place. origin stack delete-branch also moves stacked branches onto the deleted branch's parent and tells you when it checks out a different branch. See view, get, and sync.

Pull requests

  • See every owner rule on a pull request and where it stands. origin pr requirements has an Owner rules section with one line per rule, including rules that already passed: who can approve, how many files it covers, and whether it still needs an approval. Rules waiting on approval are listed first, and --json ownerRules returns the same data.
  • Clearer message when a pull request has nothing to merge. When a pull request's changes are already on its base branch, the merge blocker says so and suggests closing the pull request or pushing a commit with new changes.

Scripting

  • Pick JSON fields for origin ruleset and access commands. origin ruleset commands and the repository and namespace access commands take a field list with --json, for example origin repo access remove --json repo,action,subject. origin ruleset list --json includes bypass actors and returns an empty list when a repository has no rulesets.

Cursor mobile

Voice calls

  • Voice calls reconnect after network drops. In Cursor for iOS, a voice call that loses its connection reconnects on its own and continues the same conversation instead of hanging up. Work the agent finished while the line was down is passed along once the call is back.
  • Voice calls hang up when you go quiet. A voice call ends on its own after about 10 minutes of silence from you while the agent has no work in progress, so a forgotten call doesn't stay open. It ends like a normal hang-up, not an error, and a reply that's still playing is never cut off.

Projects

  • Project replies stream in, with the coordinator's work in Worked sections. In a Project's chat, the coordinator's replies appear as they are written, and its thinking and tool work fold into a Worked section with a duration before each reply. Chats opened from history show the same view, and a finished turn no longer leaves stray Worked rows below its final reply.
  • See when a Project is waiting on you. When a Project's coordinator or one of its workers pauses on a question or an MCP sign-in, its inbox row shows Input Needed, and the Agents pill in the Project chat reads Input Needed with a count. The waiting worker shows Input Needed in the Agents sheet instead of Working, so you no longer have to open each worker to find the question.
  • Working Projects show up in the Working tile. If a team admin has turned on Projects, the inbox Working tile on iPhone and iPad counts Projects whose coordinator or workers are running, and the Working list shows each one once with a working indicator. Archived Projects leave the Projects section and appear only in lists whose status filter includes Archived.

Pull requests

  • Restack stacked Origin pull requests. When a Cursor Origin pull request in a stack falls behind its base, the review card shows Restack instead of Update Branch, which Origin doesn't allow for stacked pull requests. Tap it to restack the whole stack, the same as on the web pull request page. If the restack hits a conflict, you see Origin's message.
  • Checks waiting on a reviewer show as Action Required. When a pull request's only failing checks are waiting on someone's approval, such as a code owner review, the status card, checks list, and agent list show a yellow Action Required status instead of failed checks, and Fix with Agent is hidden for those checks. If real CI failures are mixed in, the checks still show as failed.
  • Faster, steadier Pull Requests list. The Pull Requests sheet shows line counts and open or merged state right away from recently loaded data, then refreshes faster when an agent has many pull requests. Rows stay in place as details fill in, the Open, Draft, Merged, and Closed filters are always shown, and a message appears when a filter matches nothing.
  • Pull request links show their status in agent conversations. Pull request links in agent conversations appear as compact number chips with a GitHub or Cursor icon, matching desktop and web. When the pull request belongs to the agent, the icon shows whether it is open, draft, merged, or closed, and VoiceOver reads that status. Custom link text is kept, and copying a link still gives you the original text.
  • Tables, images, and collapsible sections render in pull request descriptions. HTML and markdown tables in pull request descriptions and agent messages, such as before and after screenshot grids, render as real tables with images and videos inside the cells instead of raw text. Screenshots attached to pull request descriptions and comments load, and the check Details sheet shows collapsible sections and line breaks instead of raw HTML.
  • Review picks up new commits automatically. With a pull request open on the Review screen, changed files and line counts update within about 30 seconds of a new commit, without pulling to refresh or reopening the app. Existing file rows stay on screen while the update loads.

Agents

  • Open files from cloud agent links. Tapping a file link in a cloud agent's transcript opens that agent's chat and shows the file at the linked line, read live from the agent's machine. Links to files in Agent Stores open too, and you get a clear message when a file is missing, isn't text, or fails to load.

Composer

  • Start agents on repos from your own machines. The composer's repo picker lists repositories checked out on your online personal machines, even when they aren't connected through source control. Pick one and the agent runs on that machine. Cloud isn't offered for these repos, so the Cloud option shows "Not available for this repository."
  • Dictation works over CarPlay and Bluetooth. Composer dictation records from CarPlay and Bluetooth hands-free microphones. If the audio route changes partway through, for example when your car connects, the recording keeps going instead of stopping.

Inbox

  • Pull to refresh the Inbox. Pull down on the Inbox on iPhone or iPad to reload your agents and the status of their open pull requests, instead of waiting for the next update or reopening the app.

iPhone Duo

  • Better support for iPhone Duo. On the unfolded inner screen of iPhone Duo, Cursor for iOS rotates with the device and stays upright when you unfold the phone while the app is open. Sheets like the model picker, search, and pull request review keep their close button in the top bar, and the Inbox profile button moves into the sidebar.

Account and billing

  • Admin Clearer plan management for team accounts. Tapping Manage Plan on a team account tells you who manages your plan: team members are asked to contact a team admin, and enterprise admins are pointed to their account team. Self-serve team admins get a Manage team subscription row in their profile that opens the team subscription on cursor.com, where the App Store region allows it. If you belong to several teams, the app opens your paid team instead of a Free one, unless you picked the Free team yourself.

Fixes

  • Voice calls. Voice calls no longer go silent when started right after dictation, show a failure message after you hang up, or play the greeting over your first words. Muting ends your turn so the assistant replies, and the live transcript no longer repeats your turn. When the microphone can't be used, the app says whether it is blocked, unavailable, or the permission prompt went unanswered.
  • Chat history and follow-ups. Earlier messages in long chats load as you scroll up instead of all at once on open, and live replies keep streaming while history loads. Follow-ups show the moment you tap Send and no longer vanish or duplicate after a reconnect, and open questions reappear when you reopen a chat. Idle Projects stay in the Projects section, and Back from a Project in the Inbox always returns to the Inbox.
  • In-app purchases and the paywall. The paywall no longer gets stuck on a blank screen or spinner; it shows an error with a Try again option instead. If a purchase you paid for was not granted because the sheet closed, buying it again recovers it instead of charging twice, and "product not found" errors happen less often. Accounts with no monthly spend limit that hit a usage block go to the web dashboard instead of a Contact Sales dead end.
  • Performance and stability. The Inbox and Composer no longer hang when you have thousands of repositories, and chats with many pull requests open faster. The app no longer crashes at launch on iOS 26 or when you rotate the device while a video is playing, and the Game of Life animated background renders again.
  • Composer, sheets, and pull request actions. The composer no longer ends up behind the keyboard after tapping Fix Checks with Agent, or drawn over the toolbar after replying to a long text selection. Search, Checks, and Pull Requests sheets have opaque headers so rows no longer show through, and the Live Activity timer hides instead of truncating in a narrow Dynamic Island. Agent and pull request links in transcripts open in the app you're using, and failed pull request actions explain what went wrong.

Agents Window

Projects

  • Search, sort, and browse a Project's Library files. In a Project's Library, All Files is now a table with Name, Kind, Modified, and Size columns. You can search files, filter by kind, folder, or modified date, sort by any column, limit the list to Project or User files, or switch to a grid of preview cards grouped by folder, and your choices are saved for each Project. Recents now update within seconds of agents writing files, and you can move through the Library with the keyboard.
  • Mention a Project's tasks from the @ menu. In a Project chat, the @ menu now has a Tasks folder that lists the Project's tasks grouped by status, in the same order as the task list in the right panel. Type a task ID (or just its number) or words from its title to find it, and pick it to add a task chip with its status icon. Mentions of Project members now show as plain @Name chips, and the coordinator's @ menu no longer offers a Branch row it can't use.
  • Faster first replies from a Project's coordinator. When you message a Project, its coordinator now replies right away with a short answer, or a quick acknowledgement and its first step, before it starts reading files or launching workers.
  • Choose which machines a cloud Project can use. In a cloud Project's chat, the Run on menu below the chat input now lists your machines and pools, so you can search and check the ones the coordinator may place workers on without asking first.
  • Add a local folder to a Project's allowed machines. In a cloud Project's Allowed machines menu, the Machines submenu now has an Add folder on This Mac row. Pick a folder and the Agents Window starts a worker for it and allows it for the Project, showing Starting and Allowing progress, with an error message if it fails.
  • Better keyboard and screen reader support for Project tasks. In a Project's task List and Board, each task's title is now the control that opens it, so Tab moves cleanly between tasks and focused Board cards scroll fully into view. Screen readers now announce each task's status, the task a status button changes, and PR counts by state such as "3 PRs: 2 open, 1 merged". Focus also follows a task when you change its status, and returns to where you were when you close find in Project notes.

Chat and composer

  • Voice calls speak the agent's own updates. On a voice call, the agent now chooses what to say: short progress notes while it works, a brief spoken answer when it's done, and a quick update when workers it started report back, instead of the call reading its chat replies aloud. Calls also show a reconnecting state with the clock running when the line drops, acknowledge tasks right away, and say so when a task can't reach the agent within 20 seconds.
  • Fixed mode slash commands dropping the prompt you typed after them. Typing a mode command followed by a prompt, such as /ask explain this file or /plan fix the tests, and pressing Enter no longer clears the input and sends nothing. The Agents Window now switches to that mode and sends your prompt.
  • Long messages you send to a Project collapse with Show more. In Project chats, long messages you send now show up to 30 lines with a fade and a Show more button that expands the full message, instead of a small scrolling box. Show less collapses it again.
  • Right-click links in Markdown files for the same actions as agent replies. Right-clicking a link in a Markdown file now shows the same menu you get on links in an agent's reply, with Open, copy, and file actions. File links open relative to the document, and links to a heading in the same file offer Copy only.

Navigation and layout

  • Split panes keep their own controls and survive reloads. When you split conversations into panes, each pane now has its own "…" chat actions menu and Project members for the chat it shows, and a tile's title bar is trimmed to just its actions menu and the show-panel button. Splits, the pane you had focused, and each tile's selected agent now come back after a reload, even when a conversation loads late or is hidden from the sidebar. A Project's own chat tab stays pinned at the front of its pane, and drag-and-drop, narrow tiles, keyboard focus, and cycling between panes behave more reliably.
  • Undo an archive with Cmd+Z, and get a warning before archiving agents with running processes. After you archive an agent or a group, press Cmd+Z (Ctrl+Z on Windows and Linux) or use Edit > Undo while no text field is focused to bring it back, and Cmd+Shift+Z to archive it again. Archiving an agent that still has background processes running now asks for confirmation first, including Archive All on a sidebar group.
  • Drop files onto an agent in the sidebar. Drag files from your desktop onto an agent or Project in the sidebar to open that chat with the files attached to its prompt. The row highlights while you drag so you can see where the files will go.

Files, changes, and pull requests

  • Easier pull request review in the PR tab. The PR tab now remembers whether you were on Changes, Description, Commits, Checks, or Reviews when you switch agents or restart, and Cmd+click (Ctrl+click on Windows and Linux) on a diff line opens that line on GitHub or Origin. Generated files start collapsed with a Generated badge, the Merge button stays disabled while an Origin rule blocks the merge, and merges that bypassed rules now show in the timeline. Pull request chips in prompts and messages show the PR's current status, such as merged or closed.

Cloud agents

  • Admin Open the live workspace for agents on self-hosted workers. Cloud agents running on a Linux self-hosted worker now show the Changes panel, file tree, terminal, and forwarded ports in the Agents Window, as Cursor-hosted cloud agents do. Start the worker with agent worker --connect-workspace start to turn it on. If the worker restarts or goes offline, the workspace reconnects when the worker is back, and an outdated worker shows a message asking you to update it.
  • Clearer status for cloud agents on self-hosted workers. When a cloud agent's self-hosted worker disconnects, the editor panel shows Worker offline and reconnects automatically once the worker is back.
  • Point environment setup at any environment.json in your repos. The environment setup dialog no longer waits to detect .cursor/environment.json. Under environment.json location, it starts on Start from scratch and lists every repo in the setup. Pick a repo to load an existing file, and enter its path, which starts filled in as .cursor/environment.json.

Browser and design mode

  • Edit CSS on an element you annotate in Design Mode. When you leave a note on a single element in Design Mode, you can open a CSS inspector to filter, edit, and live-preview its properties, then send those edits to the agent with your note. You can pin the inspector to a corner of the page while you work. Leaving Design Mode also gives the page its own cursor back right away.
  • Simpler built-in browser toolbar. The built-in browser no longer has a bookmarks bar, a favorites and history sidebar, or the Cmd+D bookmark shortcut. Your browsing history is still available from the address bar suggestions and the new tab menu.

Account

  • Admin Switch models when you hit the weekly limit. When you reach the weekly usage limit on a model your allowance does not cover, the notice can now offer a Switch Model button next to Buy Credits to move to a covered model. On Enterprise plans, the Personal Usage chart in Plan & Usage settings also opens instantly from your last saved data and shows the right spend type if you belong to more than one team.

Accessibility

  • More accessibility and reduced motion fixes. Screen readers now announce the chat transcript as a "Chat transcript" region, and announce the branch menu, project selector, font picker, and automation trigger pickers as busy while they load. With reduce motion on, dialogs fade in place instead of scaling or sliding.

Fixes

  • Steadier agent transcripts. Reopening a recent agent restores its transcript layout instead of jumping while rows re-measure. After a cloud agent reconnects, output written while you were disconnected now appears in the conversation.
  • Fixes to the subagent tray. Tray rows and breadcrumbs keep their task names and status in long chats, rows no longer get stuck loading behind one that keeps failing, and a shared Project no longer appears as its own subagent. Tracked task chips show the agent's real name instead of "Agent", and opening a finished cloud subagent reliably loads it.
  • Fixes to the prompt box and chat switching. Switching chats keeps your cursor where you left it in each prompt, and the composer glides into place instead of jumping. The fullscreen prompt now shows attached videos, stays visible but read-only when follow-ups are unavailable (for example after a failed worktree setup), and a locked prompt no longer accepts dropped or pasted files. Signing out now clears the remembered conversation, and a failed or cancelled agent creation returns you to your draft.
  • Fixes to the sidebar. Local chats open faster from the sidebar, and the window loads only your most recent chats at startup. Collapsed Pinned sections and stale sections you re-expand now stay that way across reloads, and sidebar rows no longer show stale labels while animating. Cloud agents without a repo no longer show "Cloud" twice in their tooltip, and Projects no longer appear as running when only finished workers remain.
  • Fixes to panels, menus, and the project picker. The Changes view now names nested repositories correctly and refreshes diffs when you switch repos. Opening a file that isn't on the agent's machine shows a clear message instead of a raw error, Cmd+A in the Files tree selects tree rows, and Chat actions menu picks keep focus in the dialog they open. The project picker keeps the multi-repo Save button visible while searching and stays responsive with very large repo lists.
  • Fixes to cloud Projects. In a cloud Project, a child agent's forwarded ports menu shows that child's ports, and the coordinator only plays the finished chime when it actually messages you.

Cursor IDE

Agent and chat

  • Agents stopped by an extension host crash can be resumed. If the extension host running a local agent exits mid-run, for example after running out of memory, the chat now shows an "Extension host stopped unexpectedly" error with a Resume option instead of looking like you stopped it. Affected chats are marked unread, and if Cursor is not focused, one system notification covers all agents that stopped.
  • Fixed skills not showing up when you edit an earlier message. When you edit an earlier message in agent chat, typing / now opens the skill and command picker instead of an empty list, so you can add a skill before you resubmit. The list updates when your skills, commands, or subagents change.
  • Cleaner @, #, and + menus in chat. The empty @ menu no longer lists Branch or Browser (type to find them), and a # query that matches no pull request now hides the menu, like / and @. The + menu's file item is now called Attach, Summarize appears in the / menu as soon as a chat has a message, and tooltips in menus like the mode picker no longer cover menu items.
  • Review changes with Bugbot in Mercurial repositories. Reviewing changes with Bugbot from the editor now works in Mercurial checkouts, with file paths resolved from the Mercurial root even when it sits inside a git repo or you open a subfolder.

Customization and settings

  • Admin Always Run on terminal approvals now sticks when permissions.json manages your allowlist. When your own ~/.cursor/permissions.json sets the terminal allowlist, choosing Always Run on an agent's terminal approval now adds the command to that file's terminalAllowlist, so the same command stops prompting every time. When your team's admin settings control the allowlist, the approval card no longer offers Always Run, since that choice would be ignored. See permissions.
  • Clearer MCP server status and fewer connection mix-ups. In Customize, MCP servers that are still reconnecting show "Connecting…" instead of "0 tools" or a sign-in warning, and a local server your Cursor account already has offers "Use the account version". Editing a project's mcp.json while a server connects now uses the new config, and local MCP sign-ins with OAuth now give you up to 15 minutes to finish in the browser.
  • Agent no longer gets duplicate project instructions. When CLAUDE.md is a symlink, hard link, or identical copy of AGENTS.md, the agent now receives those instructions once instead of twice. On Windows, rules, nested AGENTS.md files, and skills now apply even when paths differ only in drive-letter case.

Fixes

  • Fixes to agent reliability. Agent chats that contain large numbers now save correctly. Grep and Glob searches that time out keep the matches found so far, notebook edits keep fields the agent didn't touch, and on Windows hooks receive non-ASCII text correctly.
  • Fixes to agent chat and the prompt. Images attached in the chat prompt are no longer lost if the prompt closes while they save, and pressing send twice no longer sends a duplicate request. Questions from the agent are no longer skipped after you reopen a chat, typing in a chat moved to a separate window no longer reverses or drops characters, and Undo in the chat's file list now reverts files on Windows.
  • Fixes to the editor, terminal, and source control. Terminals you have already typed in are no longer reset when extension environment variables change, closing a terminal right after opening it no longer leaves a stray shell behind, and zsh shell integration now sets up on the first terminal. Search results group hits from the same file under one entry again, the References view now appears after Find All References, and remotes from partial clones and push-only remotes are now recognized. On macOS, an update that stalls while installing is now detected and retried.
  • Fixes to the built-in browser. PDFs opened in the built-in browser now show their pages instead of an empty viewer. Agent browser actions no longer fail right after a tab opens or moves between windows, and they stop immediately instead of hanging when the page navigates away or the browser closes.

Cursor SDK

  • Admin Choose which team an SDK agent runs under. If your login belongs to more than one team, pass teamId in Agent.create(...) options to run the agent under that team. The same option works on Agent.list, Agent.get, archive and delete calls, and on Cursor.me, Cursor.models.list, and Cursor.repositories.list. It isn't saved, so pass it again on Agent.resume and later calls. TypeScript only.
  • Claude Code hooks that deny a tool call now block it in local agents. Fixed a bug where a PreToolUse hook from a Claude Code settings.json returned permissionDecision: "deny" under hookSpecificOutput, but the local SDK agent still ran the tool call. Local agents now treat that nested response as a deny.
  • Local runs keep their output when the local agent store can't be opened. If the local database can't be opened partway through a run, a turn that already streamed text now finishes with that text as its result instead of ending as an error. The run keeps going in memory, and the stream emits a status message warning that the run's checkpoint may not be available after the process exits.

Cursor SDK

  • Python runs report why they failed. When a run fails after it starts, run.wait() in the Python SDK now returns error_code and error_message on the RunResult, and stream result events carry the same fields. Before, Python only saw an error status.
  • Fixes to local MCP servers and transcripts. Passing mcpServers on an individual send in a local run no longer drops MCP servers loaded from settingSources. Local agents also keep writing transcripts when the home directory is missing or read-only, using CURSOR_DATA_DIR when set or a private temporary directory otherwise.

Pull requests

  • See which bot owners can clear a Code Review Gate. When a Code Review Gate policy still needs approval, origin pr requirements lists any bot owner whose approval satisfies it, with the origin pr reviewer request command to ask for that bot's review. --json keeps the nativeVerdict field on gates and approvals, so scripts can tell which ones Agentic Owners decides.
  • Clearer owner approval blockers in origin pr requirements. Owner approval blockers name who must approve which paths. If the author is the only owner of the changed paths, the blocker says an approval from anyone but the author clears it. When a diff is too large for a full risk assessment, the blocker says so and suggests splitting the pull request.

Rulesets

  • Admin Ruleset bypass actors use group instead of team. In origin ruleset create and update specs and in ruleset output, a group bypass actor's actorKind is now group. Specs that still say team fail with an error telling you to use group, so update any saved specs.

Fixes

  • Fixes to update hints. The Origin CLI no longer prints its update hint or starts a background update check when git runs origin credential-helper or your shell requests completions, so git and shell output stay clean.

Pull requests

  • Clearer approval requirements. Blockers for required reviewers name the files that need the approval, for example "1 approval from X required for a.ts, 0 so far." When Agentic Owners decides a pull request's review, origin pr requirements labels it "Agentic Owners" instead of "Code Review Gate".

Repositories

  • Map GitHub authors and dry-run pull request imports. origin repo import-prs accepts --user-map <file> to map GitHub logins to Origin users (user:<email>), apps (app:<app_id>), or the import actor (-), with an optional * -> target catch-all. Add --validate to check every pull request's state, author mapping, and referenced commits without importing anything.

Rulesets

  • Admin GitHub ruleset import no longer creates bypass actors that grant nothing. Organization admin roles, repository roles, and enforce_admins are now reported as unsupported instead of becoming role bypass actors that never let anyone bypass. Ruleset commands accept only user, team, and app bypass actors, and origin ruleset update from a spec removes stored role actors. Add who may bypass with origin ruleset update.

Scripting

  • Pick origin auth status JSON fields, and catch broken git credential setup. origin auth status --json takes a field list, such as --json account,team,organization, and --json help lists every field. origin auth status also warns when git uses the Origin credential helper for some Origin remotes but not owner/repo ones, and names the origin auth setup-git command that fixes it. See authentication.
  • Pick the label fields you want in JSON. origin label list, create, edit, and delete take a comma-separated field list with --json, such as --json name,color, matching gh. Bare --json lists the fields. See labels.

Fixes

  • Admin Fixes to ruleset import and branch patterns. origin ruleset import no longer adds a require-branch-up-to-date rule when a GitHub status-check rule lists no required checks. Saving a ruleset with an invalid branch or tag pattern fails with a plain explanation of the problem and how to fix it.
  • Fixed git remote detection and origin push local checks. Commands that read your git remote recognize scp-style remotes without a user, like host:org/repo, and no longer mistake local paths containing a colon for remotes. origin push local rejects --remote origin or --remote github before any other checks run.
  • Fixes to the transcript private-key check. The Origin CLI refuses to upload a transcript that hides a private key behind a BEGIN marker split across separate strings, including when a later string or plain-text line finishes the marker. Before, these transcripts could pass the secret check.

Web

  • Admin Ask an app to break glass or grant an ownership exemption on a blocked pull request. When a pull request is blocked, its merge options now list the apps that can break glass or grant an ownership exemption. You can send a request to one of them, see it while it's pending, and cancel it. Apps grant a request by creating the exemption. See the Origin API changelog for details.
  • Admin Ownership exemptions now last through new pushes and show as waived. An ownership exemption granted on a pull request now stays in effect after you push new commits, so you don't have to grant it again. In the merge box, code owner groups and owner policies covered by the exemption now show as "Waived · exemption" with who granted it, instead of waiting or disappearing. Review requirement and mergeability responses in the Origin API also include the exemption's reason and granter.
  • Merge a pull request that targets another open pull request's branch. When a pull request's base is the branch of another open pull request it isn't stacked on, you can now merge it into that branch from the pull request page instead of being blocked. The merge box shows a Stack onto row so you can stack it on the other pull request in one click. Auto-merge still waits until the other pull request merges.
  • Updating a stacked pull request no longer brings back removed code. After a parent pull request is squash-merged, Update branch and applying a merge-conflict resolution on its child now stop and ask you to restack onto the base branch when a merge could reintroduce changes the child removed. The pull request page also hides Restack and Update branch when a restack is known to stop on a conflict at or below that pull request; merging is unaffected.
  • Commit rows credit co-authors and the person behind agent commits. Commits on a pull request now list the people named in Co-authored-by trailers after the commit author. Commits a Cursor agent made for you show your avatar with a Cursor "via" badge instead of a generic Cursor Agent avatar.
  • Fixes to reviewer groups and bots on pull request pages. A requested reviewer group on a pull request now shows its members and their review status, even when the group isn't in your reviewer list. In a pull request's Activity, review requests to a bot show the bot's name instead of a blank reviewer, and requests to a group that was later removed show the group's name instead of a raw ID.
  • Never open pull requests as drafts. A new Never open PRs as drafts setting, under Pull requests in your Origin settings, makes every pull request you or your agents open start ready for review, even when a draft is requested. It applies to all your pull requests, whether you open them from the web, the Origin API, or a cloud agent.
  • Pushes now warn you when they add a secret. When you push to an Origin repository, Origin scans the pushed changes for secrets such as API keys and tokens. If it finds one, git push prints a warning naming the commit, file, and line, with steps to remove it. The matched secret itself is never printed or logged.
  • Admin The audit log records more Origin changes. Your team's audit log now records repository ruleset creates, updates, and deletes, label changes on repositories and pull requests, merge-when-ready being turned on or off, repository deletions, review dismissals, and changes to a namespace's inbound IP allowlist. Each entry includes the acting member's public user ID, which also appears as a new User ID column in CSV exports, and app installation entries list the permissions the installation holds. Repository events that were sometimes dropped right after a repository was created or deleted are now recorded too.
  • Admin Clearer names for bypass users in ruleset settings. In a repository's Settings > Rules and Protections, people on a ruleset's bypass list no longer show up as bare user IDs. Someone you don't have permission to see reads "Hidden user", a removed account reads "Deleted user", and an account with no name or email reads "Unnamed user". Hover a row to see the reason and the user ID.
  • Admin Cloud agents work on teams with an inbound IP allowlist. Cursor cloud agents can now clone, push, use Origin MCP, and open the raw file, tarball, and clone links they create, even when your team has an inbound IP allowlist. The exemption covers only your own team's repositories. Concurrent allowlist edits are now applied one at a time, so two admins editing at once can no longer lock each other out or go over the entry limit.
  • Jump from a Cloud Agent's comment to its run. Comments, threads, and reviews a Cloud Agent posts on a private Origin pull request now end with an Open in Web button that opens the agent run that wrote them. Comments on public repositories don't include the link.
  • Slack notifications for mentions, with cleaner formatting. In Notifications settings, you can now turn on Slack DMs for when someone mentions you, and mentions written in bold or after punctuation now count. Pull request and comment bodies in Slack now show links, code, headings, lists, and collapsed sections as readable text, not raw HTML tags. When you subscribe a Slack channel to a repository, the confirmation tells you whether a test message was sent or skipped because the channel is shared with another company.

API

  • List commits that touch a file or directory through the Origin API. Listing commits through the Origin API, Origin SDK, or the Origin MCP server's list_commits tool now takes a path, so you get only the commits that change that file or directory, as git log -- <path> does. Paths containing .. are rejected. See the Origin API changelog.
  • Find pull requests by merge commit or merge time in the Origin API. You can now list the pull requests that landed as specific commits, up to 100 at a time, or those merged within a date range, so it's easier to map commits back to pull requests or build release reports. The same filters work in the Origin MCP server's pull request list tool. See the Origin API changelog for details.
  • Request, list, and remove bot reviewers through the Origin API. Requested reviewers in the Origin API now come back in a single reviewers list that includes your team's service accounts alongside users and groups, and you can request or remove any of them by kind and ID. Reviewer webhooks now fire for service-account reviewers instead of skipping them, and the Origin SDK supports the new fields. The older users and groups fields are deprecated but still accepted; see the Origin API changelog.
  • Origin MCP tools can list changed files between refs and fetch agent screenshots from pull requests. The compare_commits MCP tool has a new files view that pages through every file that differs between two refs, with per-file patches. summary is still the default. get_pull_request_attachment can also fetch cloud agent screenshots and recordings linked in a pull request, from an artifact share link or an agent's artifacts page, for users who can view that agent.
  • Admin Apps and service accounts can do more through the Origin API. Apps installed on a whole namespace can now be granted the Create repositories permission and create repositories through the Origin API. Service accounts and app installation tokens, including tokens acting as a member, can now download repository tarballs through a signed download link instead of failing. Listing namespaces now also returns namespaces other teams have shared with you. See the Origin API changelog for details.
  • Check runs reject oversized names and links, and mergeability names code owners. When you post a check run through the Origin API, a check or suite name over 255 characters or a details URL over 8,192 characters now returns an invalid-argument error instead of being accepted. When you ask the mergeability endpoint for owner rules, it now also returns the owning users' names and avatars, so you can show who owns the files instead of raw user IDs. See the Origin API changelog.

Fixes

  • Fixes to code owner reviews. Code owner review requests are no longer removed unexpectedly, and pull requests in very large stacks now get them. On pull request pages, owners named by @login now show their approvals correctly, and editing a rule prompt in an .agentic-owners directory now refreshes merge readiness. The Code Review Gate check no longer gets stuck in progress, and reporting a gate result right after a push no longer fails as a stale head.
  • Fixes to mergeability and merge conflicts. Mergeability checks no longer fail while a repository is being compacted. Pull requests with very large numbers of conflicts mark the conflicting file list as incomplete, merge results follow the repository's current .gitattributes rules, and directory-rename conflicts are now reported as their own kind of conflict.
  • Fixes to pull request pages. Pull request pages now update live when merge details or the author change, and reconnect right after the network returns. Files over 10 MiB show "Too large to count" instead of +0 -0.
  • Fixes to Slack notifications. Slack and email notifications no longer arrive twice, review reminders are no longer sent twice when clocks fall back, and reminders that don't require review requests no longer list your own pull requests. Subscribed channels no longer miss events when many channels follow a repository, and no longer get close or merge posts for drafts that were never marked ready for review. Slack reminder and channel settings now refresh after a failed save or removal.
  • Fixes to Git operations. Commits with malformed author lines or non-UTF-8 encodings now load and display as git log shows them, and changing a default branch no longer fails after an interrupted write.
  • Fixes to GitHub mirrors. Branch updates, merges, and deletes made in Origin now reliably reach the GitHub mirror, and a push no longer fails when GitHub refuses hidden refs or a branch was deleted in the meantime. Fetching pull request refs or new commits from a mirrored repository no longer times out or stalls while the mirror catches up.
  • Fixes to pull request search. An app/<slug> qualifier that can't be resolved now suggests using the app's app_… id, and searches from a repo-limited token against a repository it can't read are refused without counting against your search rate limit.

Agents

  • Admin Agents know when a subagent fell back to a different model. When a custom subagent is set to a model your team's admin settings block, that needs Max Mode, or that isn't valid, the parent agent now learns that the subagent ran on the parent's model instead. Subagents that inherit the parent's model also no longer get blocked on teams that disallow the Fast tier when the parent isn't using Fast.
  • Admin Computer use stays on when your team blocks the video review model. Fixed an issue where blocking Gemini 3.1 Pro in team model settings turned off computer use for every Cloud Agent on the team, even when the computer use models were allowed. Now only video review is unavailable, and the agent keeps testing with computer use.
  • Broken skills are flagged in the run. When a skill's SKILL.md has invalid frontmatter, no description, or can't be read, the Cloud Agent run now shows a warning naming the skill and how to fix it. Previously the agent quietly skipped the skill or was unlikely to pick it.
  • Faster follow-ups after turns that only read files. When a turn only reads or searches files, or sends messages, the next turn now starts sooner. The agent no longer waits on a workspace snapshot that wasn't needed because nothing changed.

Pull requests

  • Pull requests open ready for review by default. Cloud Agents, including Project workers, now open pull requests ready for review instead of as drafts, unless you or the task ask for a draft. A new Never open PRs as drafts switch in the Pull requests section of your codebase settings keeps every new pull request you and your agents open ready for review, even when a draft is requested.
  • Agent comments on private Origin pull requests link back to the run. When a Cloud Agent posts a comment, reply, or review on a private Origin pull request, the comment now ends with an Open in Web button that opens the agent run, so reviewers can jump straight to the agent's work.

Projects

  • Project coordinators keep the task tracker accurate. In Projects that use the task tracker, the coordinator now links each pull request a worker reports to the right task, unlinks workers from finished tasks, and closes Queued or parent tasks once other work has delivered them. Workers now report status to their coordinator.
  • Parent tasks in a Project's tracker now follow their subtasks. In a Project's task tracker, a parent task with no pull request or agent of its own now takes the status of its most active subtask (In Progress, then For Review, then Queued) and moves to done once every subtask is complete. The status updates on every change to its subtasks, so the List and Board views no longer show a parent stuck behind its children.
  • Project workers can check the task tracker. In a Project that tracks tasks, workers and child coordinators can now read the Project's task tracker to see the status of other tasks, without being able to change it. Workers that are still busy with a follow-up message, or that belong to a teammate, also no longer get unlinked from closed tasks when their turn ends.
  • Project workers get short topic names you can find later. When a Project coordinator starts workers, it now names each one for what the work is about, like "Webhook retry backoff", instead of a role, action, or PR number. A worker started from only a PR or issue number is renamed once its first report says what the work is about.

Environments

  • Manage an environment's secrets from its dashboard page. Each environment's page in the Cloud Agents dashboard now has a Secrets section where you can view that environment's secrets and add new ones, unless its configuration is locked.
  • Admin Admins can turn off personal secrets for Cloud Agents. Team admins can turn on Disable personal secrets in the Cloud Agents security settings. Members then can't add or edit personal secrets, and secrets they already added are no longer given to Cloud Agents. Members can still delete their existing personal secrets.
  • Admin Agents without a repository run on your team's shared default environment. When a member of a team starts a Cloud Agent without picking a repository or environment, the agent now runs on the team's default no-repo environment instead of a personal hidden one. Nobody can edit, rename, rebuild, delete, or add environment secrets to the team default, but personal and team secrets still apply to it. An unnamed personal environment with no repositories is locked until you give it a name.
  • New environments created by a Cloud Agent now use your repo's environment.json. When a Cloud Agent starts on a single repository with no environment yet, it now boots with the repository's committed .cursor/environment.json instead of ignoring it. The new environment keeps tracking that file, so later agents and builds use it too.
  • Agents started on a specific environment build use that build's config. Cloud Agents and subagents started on a specific environment build now boot with the environment.json and settings recorded for that build, instead of the repository's current file or the parent agent's newer environment version. This keeps an agent's setup consistent with the build it starts from.
  • Verify an environment build with an agent from the Builds section. An environment's Builds section on the Cloud Agents dashboard now has a Verify with Agent link next to Trigger build. It opens a new agent in a new tab, ready to check that the build works. The link appears when you can trigger builds and your team allows manual builds.
  • Pick Origin repositories shared with you when setting up an environment. When you create a Cloud Agents environment or edit its repositories, the repository picker now also lists Origin repositories from other teams' namespaces that you've been granted access to, after your own team's repositories. This lets you build a multi-repository environment across namespaces from the dashboard instead of only through the API.
  • Admin Tighter, more reliable restricted network access. Cloud Agents with restricted network egress no longer reach every *.cloudfront.net host by default; only the CloudFront hosts needed to pull public container images stay allowed, so add any other CloudFront domain your environment needs to your network allowlist. Repositories hosted on Cursor now clone faster under restricted egress, and changing network settings while an agent runs now applies that agent's current environment allowlist instead of a stale one.
  • Default environments no longer trust two compromised root certificates. Cloud Agents' default environments now distrust the emSign ECC Root CA - G3 and AAA Certificate Services roots after private keys for intermediates under them were exposed. The system certificate store is updated, and on the default image the Java trust store is rebuilt from it, so connections signed through these roots fail instead of being trusted.

Self-Hosted Machines

  • Open the live workspace of agents on Self-Hosted Machines. Agents running on Self-Hosted Machines can now be opened from the Cursor desktop app with their files, changes, a terminal, and forwarded ports. Turn on the live workspace when you start a Linux machine. If the machine restarts, opening the agent reconnects to it.
  • Admin Send self-hosted runs to a team default pool. Team admins can pick a Default pool in the Self-Hosted section of the Cloud Agents dashboard, and self-hosted runs that don't name a pool or machine go there instead of routing by repository alone. If the default pool serves the run's repository, it takes the run. If not, another team pool that serves the repository takes it, and an any-repo default pool takes everything else. Repository access checks still apply.

Web

  • New agent transcript on the web. Cloud agent conversations at cursor.com/agents, the pull request agent panel, and side chats now use the same transcript as the Agents Window. Your first message in a new agent shows up right away, images included, and in the agent sidebars on codebase pages the mouse wheel and Page Up and Page Down scroll the transcript.
  • Fix with agent on pull request pages goes to the right agent. On cursor.com pull request pages, Fix with agent and similar actions now follow up the agent that opened the pull request, or your most recently active agent on it, instead of whichever agent is newest.
  • Agent pages show their details sooner on the web. When you open a link to a cloud agent directly, its header, status, and starting prompt appear sooner.

Integrations

  • Linear sessions link to the agent and its PR, and PRs close their Linear issue. Cloud Agent sessions in Linear now show Open in Web, Open in Cursor, and Pull Request links, and the final reply is a plain summary that no longer appears twice. Pull requests from runs started on a Linear issue now end with a Fixes <ISSUE-ID> line, so Linear links the PR and closes the issue when it merges.

MCP

  • MCP tools that need an extra permission now ask you to grant it. When an MCP server refuses a Cloud Agent's tool call because your sign-in is missing a permission the server names, the agent shows a sign-in card that names the tool and the permission it needs, instead of failing. If signing in again can't grant it, the agent tells you an admin needs to approve it. Permissions you grant this way are kept the next time you sign in.
  • Fixes to Gmail and Google Calendar tools. When Gmail rate-limits a request, the agent now waits briefly as Google asks and retries the call, instead of having later Gmail calls blocked for a cooldown period. When a Google Calendar call fails because the connection is missing a permission, the agent now tells you to reconnect the account instead of showing a raw permission error.
  • Google Sheets calls recover when the wrong account is connected. If a Google Sheets tool call fails because the account in use can't open the spreadsheet, the agent now sees your other connected Sheets accounts and can retry with one that has access.

Fixes

  • Fixes to Self-Hosted Machines. When a Self-Hosted Machine drops mid-run, Cloud Agents now reconnect to it or move to another free machine in the pool instead of losing the run.
  • Fixes to MCP connectors. Common argument mistakes in Slack, Google Sheets, Airtable, and Jira tool calls are repaired automatically, and signing in to a team-owned MCP server no longer conflicts with your own entry for the same URL.
  • Fixes to environment builds. The Trigger build dialog now scrolls so its button stays reachable when an environment has many repos.
  • Admin Fixes to Cloud Agents dashboard errors. Team preferences on the Cloud Agents page no longer ask paid, Enterprise, or merged teams to upgrade. When a prepaid balance runs out, the error now names it instead of pointing to on-demand usage settings.
  • Fixes to agent stores. Project members can open links to a Project's store, and subagents running on their own machine keep access to their Project store after a checkpoint or finished turn. Store traffic from cloud agent machines no longer counts against your own rate limit, and the Agent Stores page shows the server's error message when it fails to load.
  • Fixes to how agents handle Origin pull requests. When an agent checks CI on an Origin pull request, checks that are still running but already failing are now flagged, so the agent can start fixing them without waiting. Merge-conflict notices say "at least N" conflicted paths when the list was cut short, and merging a pull request hosted elsewhere now points the agent to that host's tooling. Commits from subagents of a Slack thread agent now carry the same co-author line, and agents no longer commit a stray file created by the co-author git hook.
  • Fixes to app publishing. Confirming a publish no longer fails as out of date when the agent committed outside the app's folder or its machine had shut down; the machine now wakes and the confirm retries automatically. New apps are named after their folder or repository instead of a generic name, and a first publish no longer waits on missing secret values. After publishing, the agent names the repository and folder the app came from and links to it, and no longer says secret changes are waiting in the publish dialog.
  • Fixes to Pages. Published pages now show updates on the next load after you republish, unpublish, rename, or change their address. Pages no longer fail to compile when code blocks or comments mention import, and images hold their shape with a blurred preview while loading. Creating a page in a team thread where team sharing is restricted now shows your admin's sharing policy instead of a generic error.

Integrations

GitHub

  • Admin Connect GitHub just for yourself, then move it to your team. On a team's integrations page, GitHub now offers Just for me next to Connect, so team members who can't manage the team's integrations can still connect GitHub for their own use. After a personal connect, team admins see an offer on the success page to move the installation to the team, and can also move personal installations from the Details dialog on the GitHub row.
  • Admin Clearer GitHub access errors on the Integrations page. When Cursor can't reach a connected GitHub or GitHub Enterprise integration, its row on the Integrations page now shows a warning that explains the problem and how to fix it, with a Try again or Reconnect button, instead of a generic "Reconnect required." You can also disconnect GitHub when Cursor can't read its installations, instead of being stuck on an error row.
  • Clearer GitHub connection errors. If GitHub has an error or times out while Cursor checks your access to a repository, git, pull request creation, and agent tools now say GitHub didn't answer and to try again, instead of saying you don't have access. If your GitHub connection was revoked, Cursor asks you to connect GitHub again instead of showing an internal error. If GitHub won't refresh the connection because your primary email isn't verified, Cursor tells you to verify it.
  • Agents recover from common GitHub errors. When an agent works with GitHub, short rate limits are waited out and retried automatically, and merging a pull request without naming a method uses a method the repository allows. Merging an already-merged pull request, opening a pull request that already exists, or creating a branch that's already at the right commit now succeeds with a note. Errors that remain name the fix, such as a missing branch or label, a pull request with conflicts or blocked by branch protection, or a repository where the account only has read access.
  • GitHub pull request lookup no longer fails on busy branches. Finding the pull requests for a GitHub branch with more than 100 pull requests now returns the newest ones instead of failing, and lookups for other branches in the same request still go through.

Bitbucket

  • Admin Bitbucket Data Center works on hosts without account linking. On a Bitbucket Data Center host where members can't link their own accounts, Cursor now accepts pull request links from repositories your team has connected, and team admins' runs started from Slack, Microsoft Teams, or Jira use the host's connection. Other members, and anyone who comments @cursor on a pull request, now get a reply saying Cursor can't verify their account and asking a team admin to set up account linking, instead of a refusal or no response.
  • Agent commits on Bitbucket Data Center credit you. When a cloud agent you launch commits to a Bitbucket Data Center repository, the commit is authored as your linked Data Center account, so Bitbucket attributes it to you. If your account isn't linked or no longer has access to the repository, commits keep the default Cursor agent author.
  • Admin Switch your team's Bitbucket workspace from the dashboard. If you installed the Cursor app in another Bitbucket Cloud workspace, the Bitbucket Cloud integration now shows it and lets admins switch to it from Switch workspace in the Manage menu. A team uses one workspace at a time, and linking a second workspace now explains how to switch instead of failing with a generic error.
  • Fixes to Bitbucket Data Center pull requests. Cursor now shows diffs for Bitbucket Data Center pull requests, finds a pull request by its source branch, and opens create pull request links on your server's browser URL. Merges use a strategy the repository allows, and pull requests that can be merged no longer show as blocked by branch rules. When creating or merging a pull request fails, Cursor explains why, for example that the branch hasn't been pushed, there are no new commits, or a pull request is already open.
  • Fixes to Bitbucket Data Center. Repositories are now recognized as connected when the URL uses different capitalization, or when the instance is registered on a non-standard port. Agents can clone same-host submodules listed by their /scm/ clone URL, and PR creation no longer fails while a rejected member token stays cached. Members whose Bitbucket user was deleted are asked to reconnect instead of hitting repeated temporary errors.

GitLab

  • GitLab merges run as your GitLab account. Merging a GitLab merge request from Cursor, or turning auto-merge on or off, now uses your connected GitLab account, so protected-branch merge rules apply to your role instead of failing with an unauthenticated error. If GitLab denies the merge, the error says whether to ask a project Maintainer for merge access or to connect your GitLab account to Cursor.
  • Fixes to GitLab. Merge request state, author, and mergeability now update right away after merging, closing, reopening, or renaming instead of showing stale values. Expired GitLab OAuth tokens are refreshed automatically during permission checks, service account tokens no longer rotate repeatedly and break access, and rate limits are reported as rate limits rather than project access errors. Repository access checks also retry after a timeout or dropped connection.

Dashboard and admin

Security and audit logs

  • Admin Review and revoke your members' personal API keys. Enterprise admins can open the new Members tab on the API keys page to see members' personal Cursor API keys, filter by member, and revoke one key or all of a member's keys. The Members page also adds a View API keys action for each member.
  • Admin See which API key acted in the audit log. When an action was made with an API key, the audit log now shows the key under the user, such as a service-account key or Admin API key with its name and ID, plus the delegated user when there is one. Audit log CSV exports add Actor API Key Kind, Actor API Key ID, Actor API Key Name, and Actor Delegated User Email columns. See audit logs.
  • Admin More Origin activity in the audit log. On Enterprise teams using Origin, the audit log now records repository ruleset changes, repository deletions, labels, merge when ready being turned on or off, dismissed reviews, and ownership exemptions. It also logs app signing keys being added or revoked, and members adding or deleting personal SSH keys, with the key name, fingerprint, and algorithm. You can pick each new event in the event type filter.

Deployed agents

  • Admin Control who can view, deploy, and manage each deployed agent. Each deployed agent now has Viewer, Editor, and Owner roles: Viewers see the agent and its logs, Editors can also deploy and stop it, and Owners manage secrets, access, the Slack app, token rotation, and deletion. New agents make the deployer Owner and the rest of the team Viewer, and existing agents keep their current access. The deployed agents page also shows your deployment count against your team's limit, and deploys over a limit are refused with a message explaining which one.

Agent settings

  • Admin Turn off personal secrets for Cloud Agents. Team admins can turn on Disable personal secrets in Cloud Agents in the Cloud Agents security settings. Members then see Add and Edit for personal secrets disabled, with a note explaining why, and can still delete secrets they already added.
  • Admin Pick a default pool for self-hosted Cloud Agents. Team admins can now set a Default pool in the self-hosted machine settings of the Cloud Agents dashboard, so self-hosted Cloud Agents that don't name a pool or machine run there. If the default pool doesn't cover a run's repository, that repository keeps using its own pool. Choose None to keep routing as before. See Self-hosted pools.

Plugins and MCP

  • Manage your personal plugins and MCP servers from a team workspace. In a team workspace, the Plugins and MCP pages now have a Team and Personal switch, so you can view and manage your own plugins and MCP servers without leaving the team. Your MCP servers stay listed even when team policy blocks them, marked "Disabled by admin", and adding or editing your own servers doesn't need team manage permission.
  • Admin Import a team marketplace alongside your default one. On the Plugins & MCPs page, teams without Enterprise can now import a team marketplace even when the team has a default marketplace, because the default no longer counts toward the one-marketplace limit.

Integrations

  • Admin Connect GitHub just for yourself on a team. On a team's Integrations page, GitHub now offers Just for me next to Connect, so members who aren't admins can connect their own GitHub account. Details on a connected GitHub row lists who owns each GitHub account, and team admins can choose Move to this team to turn a personal connection into a team one. Inactive teams are warned that this needs active seats. See GitHub.
  • Source control connection problems now explain how to fix them. When a GitHub, GitHub Enterprise, or GitLab connection on the Integrations page runs into an access problem, such as a rate limit or expired authorization, the row now says what went wrong and how to fix it, with a Try again or Reconnect action. Environment build failures caused by source control errors use the same wording and link to Integrations, the docs, or the installation settings. A GitHub connection whose installations fail to load still shows as connected and offers Disconnect.
  • Admin Switch the Bitbucket workspace your team uses. If you install the Cursor app in a second Bitbucket workspace, the Bitbucket integration on the Integrations page now says so and offers to switch the team to it. Admins can also choose Switch workspace from the Manage menu, which explains how to move the team to another workspace. Members who aren't admins are told to ask one.

Cloud agent environments

  • Create Cloud Agent environments without a repository. In the dashboard's Environments section and the New environment dialog, you can now skip picking a repository, name the environment, and save it. Repository-free environments then appear in your environment lists and can be configured like any other, with an install script and secrets. See environment management.
  • Set how long a cloud agent environment's install script can run. Environment Build Settings in the dashboard now include an Install Timeout field, so a slow install no longer fails at the fixed default. Enter a whole number of minutes from 5 to 300, or leave it blank to keep the 90-minute default. When an install times out, the build failure message now points you to this setting. See managing Builds.
  • Admin Refreshed secrets tables and safer bulk scope edits. Team Secrets, My Secrets, and the Secrets section on an environment's page now share one secrets table, with new dialogs for adding and editing secrets. When you bulk-edit the repository scope of secrets with different scopes, the picker now starts empty instead of defaulting to all repositories.

Members and groups

  • Admin Set spend limits on automation and agent service accounts. Admins can now use Edit monthly spend limit on the service accounts that run automations and cloud agents, so these accounts get a budget like any other. Key, suspend, and archive actions stay unavailable for these managed accounts.
  • Admin Every member can see whether a group is Synced or Manual, and group rosters sort by usage. The Type column on the Groups tab now shows Synced or Manual for every viewer, not only for admins who manage SCIM. On a group's member list, you can click the usage column header (monthly spend, included spend, or requests) to sort by it if you can see member spend.

Bugbot

  • Admin Choose whether Bugbot reviews outside contributors on public repositories. A new Public repositories switch in Bugbot team settings controls whether Bugbot automatically reviews pull requests from people outside your team on public repositories. It starts on for existing teams and off for new teams. See Bugbot configuration.

Account and navigation

  • Old settings links now open the right place in Settings. Bookmarks and links to the old My Settings and Voice pages now open the matching section of the Settings dialog, including Pull requests preferences. If you're not on a team, a model consent link from the editor opens Settings > Privacy with that model's consent dialog already open.

Fixes

  • Fixes to deployed agents. A deployed agent's page no longer fails to load while its Slack app is waiting for Slack's first request, and now shows that it is pending. Running Slack setup again from the dashboard keeps any extra bot scopes the agent's existing Slack apps already use. The Source value and the egress allowlist domains are now easier to read.
  • Admin Fixes to the Apps pages. When an admin changes access on a teammate's app, the Private and Team options now name the app's owner instead of saying "you." Failed Revoke, Change access, and Unpublish actions now show the reason they failed instead of a raw error code.
  • Admin Fixes to MCP, LLM Gateway, and plugin marketplace settings. Editing an HTTP MCP server no longer drops its custom CA bundle when you save other changes, and long MCP tool names truncate in a single alphabetical list instead of scrolling sideways. The LLM Gateway model test lists your team's published models and says when model discovery returned nothing, instead of showing a misleading provider mismatch. Failed plugin marketplace refreshes now explain the cause, such as missing repository access or rate limiting.
  • Fixes to cloud agent environments. The Trigger build dialog on the Environments page no longer grows past the screen for environments with many repositories. Its branch pickers now scroll inside the dialog.
  • Admin Fixes to usage and analytics. The analytics overview now marks the chats and models widgets as unavailable when part of the date range fails to load, instead of showing those days as zero. On the Usage page, owners of a personal team now see their usage from before the team was created, and chart tooltips show a total for the day. Paginated tables format large counts with thousands separators.
  • Admin Fixes to members and groups. Members with an unlimited spending limit now show a blank in CSV exports instead of a very large dollar amount, and they sort last with other no-limit members. The suggested Increase limit amount no longer jumps to a huge value because of them. A group's Codebase access section now shows that the team has no Codebase namespace instead of failing.
  • Admin Fixes to team invites and the team list. Sending invites or getting an invite link no longer times out on teams with very large numbers of pending invites, and pending invites waiting on a card confirmation now say so, can be revoked, and complete correctly for shared invite links. The Invite members dialog no longer offers a Free seat on teams that can't use one, and the team list no longer shows empty teams left over from abandoned checkouts.
  • Admin Fixes to audit log labels. Source control connection entries in the audit log now show "Bitbucket Cloud" and "Bitbucket Data Center" instead of raw provider IDs. The API keys page and audit logs now label Grok inference keys as "Grok API" keys.
  • Fixes to the account menu, profile, and sign-in. Approving a desktop sign-in now tolerates a slow security check, and the approve button can't be pressed twice. Accounts that signed in by phone no longer see a placeholder email address, and picking or removing a profile photo in Settings now saves right away. The account menu shows your profile photo and lists the same items everywhere.
  • Fixes to dashboard navigation and the Overview page. Switching teams from the team switcher or account menu no longer lands on a "team not found" page, and older dashboard links now keep their section anchor and open the matching Cloud Agents view and team instead of the dashboard home. On the Overview page, clicking the Cursor card opens the desktop app when it's installed and falls back to the download page otherwise.
  • Fixes to automations. Breadcrumbs, Back links, and navigation after deleting an automation now return you to the Mine or Team list you opened it from. The model picker in an automation's prompts section is now searchable.
  • Fixes to dialogs and menus. Pressing Escape or clicking outside a menu inside a dialog now closes just the menu, and menus in side panels are no longer cut off at the panel edge. The Duplicate as Origin dialog now fits on screen, and its list scrolls.
  • Fixes to scrolling in the Changes agent panel. The agent conversation in the dashboard's Changes agent panel scrolls with the mouse wheel and Page Up/Page Down again, instead of only responding to dragging the scrollbar.
  • Fixes to merging accounts. When a merge can't go ahead, the dashboard now says why, for example a merge already in progress or a team or organization that blocks it. The merge banner on Overview appears only when there is an account you can merge, and Profile settings no longer shift while it loads. Merged accounts on a merged team now open the new dashboard instead of the legacy one.

Automations

Triggers

  • Security Reviewer runs on GitHub label changes. Security Reviewer automations with a Pull request label changed trigger now run when a matching label is added to or removed from a GitHub pull request. Before, they never ran. To apply the fix to an existing automation, open it and save it.

Automation runs

  • Admin Automations start on teams with an inbound IP allowlist. Personal automations on teams with an inbound IP allowlist no longer fail with an "IP address is not on this team's inbound IP allowlist" error. Scheduled runs, webhook deliveries, and CI triggers start normally, and Cursor still checks your sign-in and repository access. Manual runs and API-key webhook calls still have to come from an allowed IP address.
  • Clearer failure messages for runs and webhooks. When a webhook-triggered automation fails to start, the webhook response gives the same failure title and reason as run history, instead of a generic error. Rate-limited runs explain that the limit can apply across all of Cursor and suggest trying again after a short wait. Runs whose agent fails to start tell you to retry and to contact support if it keeps happening.
  • Scheduled automations start up to 5 minutes after their set time. Each scheduled run can start up to a fifth of its interval late, capped at 5 minutes. This applies to every schedule, not only hourly and top-of-hour ones. A schedule that runs every 15 minutes can start up to 3 minutes late, and a daily schedule up to 5 minutes late. How often a schedule runs does not change.

Integrations

  • Clearer errors when connecting Linear or linking Jira. If you aren't a team admin, connecting Linear from a Linear trigger tells you only team admins can connect it, instead of showing a generic failure. Linking your Jira account explains when you aren't on the Cursor team connected to that Jira site, or when the site isn't connected to a Cursor team yet.

Fixes

  • Fixes to the repository picker. The repository picker in the automation editor labels partial lists as Recent, shows a spinner while more repositories load, and says clearly when a search fails.
  • Fixes to the PR Approver and pull request creation. The PR Approver automation finds routing policies in .cursor/approval-policies/, including ROUTING.md, and reviewer-specific criteria in .cursor/approvals/. Automations that open GitHub pull requests no longer fail silently when Cursor can't find the repository. The agent gets a clear error instead.

APIs

  • Admin Team admins can list a member's personal environments. The Cloud Agents API's GET /v1/environments accepts a userId query parameter. With a team admin's user API key, pass a member's ID to see their personal environments that list at least one repository, all of which you can read, for example to promote them to team environments. Pages can come back short or empty, so keep requesting until nextCursor is absent.
  • Admin Match team directory groups to the group IDs used elsewhere. Team directory groups in the Admin API now return each group's publicId (grp_…), the ID access grants use for the group. GET /teams/directory-groups/:groupId and its /members route also accept the publicId; write routes still take the team_group_… ID.

Security and approval agents

Security agents

  • More accurate fix detection for earlier findings. When you push to a PR, Security Reviewer now recognizes a fix made outside the flagged file, such as a new check in a caller elsewhere in the PR, and closes the finding. It closes a finding only when the fix appears in the lines your PR changes, and it rechecks every earlier finding before posting results.
  • Security reviews no longer fail at startup when the model list is briefly unavailable. If the model list is rate-limited or temporarily unavailable, the review retries with backoff, honoring the server's requested wait, before giving up.

Approval agents

  • Fixed approvals held by a missing rule rating. Fixed an issue where PR Routing & Approval could hold a PR's approval because a matching rule, or its global override copy, never got a risk rating. A rule and its override copy now share one rating, based on the files that rule governs.

Team settings

  • Admin Choose whether Bugbot automatically reviews outside contributors on public repositories. Admins can turn off automatic reviews of pull requests from bots and people outside the team on public repositories. Use the Public repositories setting in Bugbot's Team settings. Manual reviews, private repositories, and pull requests from team members work as before. Existing teams keep reviewing these pull requests until an admin changes the setting, and new teams start with it off.
  • Admin Fixed Bugbot skipping reviews for some Enterprise teams. Bugbot no longer skips pull request reviews with "Bugbot plan is disabled for this team" on Enterprise teams whose dashboard shows Bugbot as enabled. On Teams plans, Bugbot still follows the team's Bugbot setting.

Reviews and checks

  • Admin Bugbot checks no longer stay pending when your team hits its daily cap for bot pull requests. When your team reaches its daily cap of reviews on pull requests opened by bot accounts, Bugbot finishes its check right away and shows the cap message. On GitHub, Bitbucket, and Azure DevOps the check fails, so a required Bugbot check blocks the merge until Bugbot reviews the pull request. On GitLab it finishes as neutral.
  • Bugbot explains when it can't push Autofix changes. When you ask Bugbot to push Autofix changes and the push fails, Bugbot replies on the pull request with the reason instead of staying silent. The reply names the cause, such as a merge conflict, branch protection, missing permission, or a mirrored repository.

Fixes

  • Fixes to GitLab webhooks. GitLab no longer disables the Bugbot webhook after repeated errors. Bugbot skips merge request events that arrive without a head commit instead of failing on them. When Bugbot rejects a webhook delivery, GitLab's recent events show how to fix it: delete leftover webhooks in the project's Settings > Webhooks, and turn Bugbot off and on for the project if reviews stopped.

Cursor CLI

Hooks and subagents

  • Post-tool hooks now fire for subagents. postToolUse and postToolUseFailure hooks that match the Task tool now run when a subagent finishes or fails. Before, only preToolUse fired for subagents. Any additional_context a hook returns now reaches the main agent, and Claude Code PostToolUse matchers for Task work too.
  • Hooks in subagents get the right transcript. Subagent transcripts are now saved inside the parent conversation's folder, under agent-transcripts/<parent>/subagents/. Hooks that run inside a subagent now receive that file as transcript_path and CURSOR_TRANSCRIPT_PATH.

Self-hosted workers

  • Health checks track self-hosted workers through shutdown. On a worker started with --management-addr, /readyz now keeps answering while the worker shuts down, returning 503 with draining: true so orchestrators stop sending it work. The worker now fails to start if it can't bind the management address after a few retries, instead of running without health checks. Pool workers also exit after dropping a stale claim, so your host can replace them.
  • Fixes to self-hosted workers. Workers running agent worker no longer drop their connection when heavy local builds stall the machine for a few seconds, and no longer exit mid-session when briefly offered another agent. Closing the terminal or losing the SSH session now shuts the worker down cleanly, and a worker that is shutting down refuses new runs instead of accepting one it can't finish.

Agent

  • Fixed unanswered approvals being reported as cancelled. If you send a new message while a tool call is still waiting for your approval, the agent no longer tells you that you interrupted or cancelled it. Tool calls that already finished before a pause also no longer run a second time when the agent resumes.

Rules

  • Fixes to rule matching on Windows. Nested AGENTS.md files, rules, and skills now apply on Windows even when the workspace path and rule path differ only in drive-letter case.

Cursor SDK

  • Fixed local agents failing on models that only run in max mode. Local agents created with a model variant that requires max mode, such as a 500k-context selection, now send max mode automatically instead of failing with invalid parameters, and keep it across later sends and resumes. Cursor.models.list() marks those variants with isMaxMode, and ModelSelection accepts an optional maxMode.
  • Computer use tool calls show up in run events. When an agent uses computer use, tool call events now carry a typed computerUse call instead of dropping it. args.actions lists each action (click, type, scroll, key, drag, screenshot, and more), and result reports the action count, duration, screen settle time, and an optional screenshot path or error message.
  • zod is now a peer dependency. @cursor/sdk now uses the zod you install yourself (3.25 or later in the 3.x line, or 4.x), so your schemas and the SDK's share one copy. The exported schemas are built with the zod/v3 API, which both supported versions include. Add zod to your dependencies if your project doesn't already have it.
  • Slow-starting stdio MCP servers can take longer to connect. Local agents now honor startupTimeoutSec on a stdio server in your MCP config, so a server that is slow to start no longer fails at the fixed 60-second connect limit. You can set up to 5 minutes; servers without the setting still default to 60 seconds.
  • Brief server outages no longer fail one-off SDK requests. When the server reports a temporary outage that is safe to retry, the SDK resends a one-off request once, after a short random delay of up to 1 second, instead of failing right away. Streaming calls such as agent runs are never replayed.
  • Background subagents keep running after the tool call that started them. When a local agent starts a subagent in the background, the subagent no longer gets cancelled as soon as the launching tool call returns.
  • Fixed sandboxed runs failing on a non-executable sandbox helper. The SDK's platform packages now install the sandbox helper with execute permission.
  • Fixed Python cloud launches with a named cloud environment. Launching a cloud agent from the Python SDK with a named cloud environment and no repos no longer fails validation.

Pull requests

  • Safer transcript uploads in origin pr gate report. --transcript-file refuses a transcript when a private key's BEGIN marker is split across lines, or when a line is nested too deeply to check for secrets. The help text and the "Stored transcript" message say that anyone who can read the repository can read the transcript until 30 days after the pull request closes or merges.

Rulesets

  • Admin Clearer guidance when importing GitHub team reviewer rules. origin ruleset import --from-github lists each required reviewer team with its approval count and tells you to add the requirement with origin ruleset update using requiredReviewers. It also notes when GitHub file patterns can't narrow the requirement, since Origin applies it to every pull request on the targeted branches. See rulesets.
  • Admin GitHub file size limits carry over on ruleset import. origin ruleset import --from-github converts a GitHub push ruleset's maximum file size rule into an Origin push ruleset instead of dropping it. Limits from 1 to 100 MB carry over; other values are still reported as unsupported.

Authentication

  • Origin App installation tokens last their full lifetime. When you export an Origin App installation token as CURSOR_AUTH_TOKEN, origin uses it until it expires instead of rejecting it five minutes early. Once it has expired, the CLI tells you to mint a new one.

Scripting

  • Clearer errors for file and stdin flags. Flags that take a file or - for stdin, such as --input, --body-file, --signing-key, and --public-key, require a value and fail with a usage error that names the flag and the problem, such as "does not exist" or "is not valid JSON". origin pr create --body-file - no longer competes with the push prompt for stdin.

Agents Window

Browser and design mode

  • Edit CSS in Design Mode more like browser DevTools. The CSS inspector in Design Mode notes now works like browser DevTools. Use the Up and Down arrows to step a value (hold Shift for steps of 10), drag a number to scrub it, or type a bare number like 8 to get 8px. Color and length rows suggest the page's own CSS variables, and the values shown now come from the browser's real cascade, including nested rules, adopted stylesheets, and cross-origin stylesheets.
  • Pick parent elements with the arrow keys in Design Mode. While hovering in Design Mode, press Up to move the highlight to the parent element and Down to go back, then click to select it. The hint bar shows the shortcut, and the selection ring around a picked text element now sits slightly off the text instead of hugging the letters.
  • Browser tabs reload on their own when your dev server comes back up. A browser tab pointed at a local port that had stopped responding now reloads when a terminal command or an agent's background shell starts serving that port again, so you no longer have to refresh by hand after restarting a dev server. A new Reload Browser Tab command (Cmd/Ctrl+R) reloads the active tab.

Navigation and layout

  • A cleaner, faster command palette. The Cmd+K palette now offers Select Model, Archive, and New Side Chat. Paste a pull request URL or type a pull request number to jump straight to that pull request. Rows fit on one line, shortcuts shown next to commands match what runs them, and the palette opens and searches faster when you have many agents. Ctrl+Tab now opens the recent agents switcher, which uses the same look as the palette, even while the palette is open.
  • Back reopens tabs you just closed. After closing a tab in the Agents Window, such as a file, terminal, browser, Changes, Output, or a side chat, press Back to bring it back where it was, with its navigation history intact. Cmd+Shift+T focuses a tab that Back already restored instead of opening a duplicate. Mouse side buttons pressed in a chat pop-out window also no longer navigate the main window.

Files, changes, and pull requests

  • Play audio files and scroll through whole PDFs in file tabs. Opening an audio file such as .mp3, .wav, .ogg, .aac, or .opus in a file tab now shows a built-in player instead of a binary placeholder, and it pauses when you switch away. PDFs now scroll continuously through every page, the toolbar shows the current page with buttons to jump between pages, and zooming keeps your place. The CSV preview's row-number column also no longer lets scrolled cells show through.

Cloud agents

  • Agents on Self-Hosted Machines are grouped by machine in the sidebar. The sidebar now groups agents running on Self-Hosted Machines into one section per machine or pool, labeled with the machine name even before a worker picks up the run. Agents keep the machine you picked across refreshes, forks, and other windows, and for workers that can't serve live changes, the Changes tab explains why instead of spinning.
  • Add a folder to a Project's allowed machines. In a Project's Allowed machines menu, the Machines list now ends with an Add folder row for this computer. Pick a folder and the Agents Window starts a worker for it and allows it on the Project, showing Starting and then Allowing until it's done.
  • More reliable forwarded ports for cloud agents. In the editor panel's forwarded ports menu, a cloud agent port that can't be reached now shows a red dot and the error, such as a timed-out connection, until a connection succeeds. Opening a second cloud workspace no longer breaks port forwarding for the first, and turning off Auto-Forward Ports now closes ports that were forwarded automatically and keeps them from being forwarded again.
  • Fixed wrong Dockerfile path errors in cloud environment config. Editing environment.json now checks build.dockerfile and build.context against the paths the cloud build uses, resolved from the repository's .cursor folder. False missing-Dockerfile errors go away, a missing build context folder is flagged, and agents setting up an environment write paths that build correctly.

Customization and settings

  • Admin Claude Opus 5.5 and Sonnet 5.5 on AWS Bedrock. Teams that route models through AWS Bedrock can now pick Claude Opus 5.5 and Claude Sonnet 5.5 in the Agents Window's model settings.

Privacy and security

  • Safer handling of branch names in git and GitHub commands. Fixed an issue where a branch name or commit ref containing shell characters could run unintended commands during pull request lookup, CI reruns, or the branch diff. Unsafe names now skip the GitHub CLI lookup, and the diff falls back to comparing against origin/main.

Agent and chat

  • Add links and media from the transcript to chat. Right-click a pull request, agent, project task, canvas, file, folder, or line-range link in a conversation and choose Add to Chat to insert it as a chip, or as a plain link for other URLs. Images, videos, and text artifact cards also offer Add to Chat alongside Download, Copy Image, or Copy Text, and web search results, terminal citations, and thinking blocks now open the same link menus.
  • Links added to chat become compact chips. Using Add to Chat on a GitHub issue, commit, Linear issue, or web link in an agent reply now inserts a short labeled chip with an icon instead of the raw URL. The full URL is still sent to the agent and shown on hover, and clicking the chip opens the link.
  • Inline code formatting in the prompt. Text you type or paste between single backticks in the prompt now shows as inline code, and is sent with its backticks. Press Backspace right after the conversion to get the plain backticks back.
  • Find any pull request from the # menu by number. In the prompt box's # menu, typing a pull request number that isn't in the list now looks it up and adds it, with a loading status while it searches and "No matching pull requests" when nothing matches. Pull requests you opened moments ago on Origin also show up in the menu within seconds.
  • Clearer live status while agents work. When a subagent needs your approval or an answer, the main conversation's status line now says "Waiting for approval" or "Waiting for response" and shows what it's asking for. Each finished turn shows how long the agent worked, the transcript stops jumping when a turn ends, and messages sent by another agent name the agent that sent them.
  • Cmd/Ctrl+F works inside canvases. Pressing Cmd/Ctrl+F in a canvas opens its find bar, or puts focus back in the bar if it is already open, without losing your current match.

Fixes

  • Fixes to Annotate. Escape now exits Annotate (or dismisses a model warning) without collapsing the apps panel or dropping your note, and dictation goes into a focused note. Element and region screenshots frame the right area on zoomed tabs, scrolling a note or the CSS inspector no longer scrolls the page, and pin hover cards no longer block picking elements beneath them. Notes no longer jitter as they resize, the annotate cursor no longer lingers after you leave, and clicking a saved note while drafting a comment opens it after asking whether to discard your draft.
  • Fixes to switching between agents and keyboard focus. Going Back to a subagent no longer clears Forward history, and Cmd-N reliably puts the cursor in the agent's prompt without jumping mid-typing. Shortcuts no longer reach hidden agents, the hidden chat behind fullscreen Apps, or the main window from a chat pop-out. Files and pull requests opened right after starting a New Agent now stay open.
  • Fixes to trays above and below the prompt. Notices on the new-agent screen now show one at a time without shifting the composer, and pressing Escape on a blocking tray no longer opens the Browser pane. Trays in short agent tiles stay inside the tile instead of overlapping the tabs, focus returns to the prompt when a tray you were typing in gets covered, and trays you opened or closed stay that way when you switch agents.
  • Fixes to files and the editor panel. Opening a file path that doesn't exist yet from the command line now opens a new editor that saves to that path, instead of a tab that fails to load. On macOS, background file watching and scans no longer reach into protected folders like Desktop, Documents, and Downloads when your workspace contains them, so you stop getting access prompts. While a workspace connects, the editor panel shows a small spinner: hover it to see the status, or click it to view the remote connection output.
  • Fixes to window appearance. On macOS, the window's translucent background tint now updates when you switch between light and dark themes. The collapsed apps rail beside a conversation is no longer faded out when you scroll the transcript.
  • Fixes to outage notices. The sidebar status card now shows the real incident name and latest update from the Cursor status page, and says when the incident is resolved. The Cloud option in the Run on picker only reflects Cloud Agents incidents, without the extra warning text.
  • Fixes to the Cloud Desktop tab. The Cloud Desktop tab draws screen updates more efficiently and stops streaming while another tab is selected. When you switch back, it reconnects reliably with a fresh connection instead of one that may have expired.
  • Fixes to voice calls. Voice call transcripts with cloud agents now sync while the call is live, so other devices show turns mid-call. Call cards for calls on another device no longer flash an inflated duration, and the return bar in other agents shows the called agent's current title after a rename. Closing a window while voice was starting no longer crashes the voice service.
  • Fixes to agent transcripts. Messages sent from another device no longer remount or land out of place in the transcript. Cloud agent output written while disconnected now appears after reconnecting instead of going missing.
  • Fixes to the prompt box menus. The / menu no longer lags when many commands match, and typing after choosing /add-plugin no longer freezes the input. Arrow keys now wrap around in the @ and + menus, right-clicking inside an open menu no longer opens a second menu, and tiny attached images now preview at a readable size.
  • Fixes to the @ mention menu. Late-arriving results no longer move the row you've highlighted, rows no longer flash while results load, and slow sources still show up. For cloud agents, the menu shows the agent's last known files right away, refreshes on its own when the file list loads, and only suggests folders the agent's machine has.
  • Fixes to File Blame in multi-folder workspaces. File Blame now reads history from the repository that owns the file when the workspace has several folders, instead of always using the first one.
  • Fixes to reviewing changes. Opening a large GitHub pull request for review no longer freezes the window. Reviewing an agent's changes on long chats that touched many or large files uses less memory and stays responsive.
  • Fixes to blocked pull request merges. The merge button in the diff header now says why a merge is blocked, showing "Needs restack" when the branch must be restacked onto its base, or "Checking for conflicts…" while the conflict check is still running.

Cursor IDE

Cloud agents

  • Cloud agents on Self-Hosted Machines stay on the machine you picked. Cloud agents started on a specific Self-Hosted Machine keep that machine after the agent list refreshes and when you fork them, so follow-ups run on the same machine.
  • More reliable forwarded ports for cloud agents and remote windows. Connections to a cloud agent's forwarded ports that get no answer now close after 20 seconds instead of hanging and blocking other ports, and opening a second cloud workspace no longer breaks port forwarding for the first. With remote.autoForwardPorts off, ports that were forwarded automatically are no longer restored when you reopen a remote window.
  • Fixed wrong Dockerfile path errors in cloud agent environment files. Editing .cursor/environment.json now checks build.dockerfile and build.context against the same paths the cloud build uses, resolved from the repository's .cursor folder. False "Dockerfile does not exist" errors go away, a missing build context folder is now flagged, and hover text explains the path rules. See cloud agent setup.

Customization and settings

  • Agent Store sync keeps working after you switch teams. Your local Agent Store no longer stops syncing when you change your active team, as long as you're still a member of the team the store was created under. When sync is briefly refused, Cursor now waits the suggested interval before retrying.
  • Skill warnings and clearer skill menus. Customize > Skills now shows a warning icon on skills whose name, description, or body go past the Agent Skills limits; hover to read what to fix, or click to copy it. Skills that share a name show as separate rows in the / menu instead of collapsing into one, skills appear there sooner after you open a window, and /skill mentions stay highlighted in sent messages.
  • Delete Old Chats reclaims disk space. After you run Developer: Delete Old Chats, Cursor offers Rewrite Now to give freed space in its local database back to your disk. Cleanup also removes leftover message data from deleted chats without freezing the editor, and deleted chats, including ones moved from a cloud agent to local, no longer reappear in the chat list.
  • YAML keys stand out in the Cursor Dark themes. In Cursor Dark, Cursor Dark High Contrast, and Cursor Dark Colorblind, YAML keys now have their own color instead of matching their values. Booleans and null now use the regular text color, so paused: false no longer looks like a string.
  • View in Cursor opens the plugin you were looking at. Clicking View in Cursor on a plugin's marketplace page now opens that plugin's page in Customize, instead of the general plugins settings.

Agent and chat

  • New installs steer the running agent by default. On a fresh install, messages sent while the agent is running, including ones sent manually from the queue, now steer it instead of queuing or interrupting, and Steer is listed first in Settings. Existing installs keep their current behavior.
  • Add file links from chat back into the composer. Right-click a file link in an agent reply and choose Add to Chat to attach that file to your current message. File and link menus now also work in thinking blocks and on file chips in your own messages, and clicking a chat mention in a sent message opens that chat.
  • Inline code formatting in the agent prompt. Text you type or paste between single backticks in the agent input now shows as inline code, and is sent with its backticks. Press Backspace right after the conversion to get the plain backticks back.
  • Clearer live status in agent chat. When a subagent needs your approval or an answer, the main chat's status line now says "Waiting for approval" or "Waiting for response" instead of "Waiting for subagent". The transcript also stops jumping when a response finishes, and messages sent by another agent name the agent that sent them.
  • Background shells keep running between turns, and archiving a chat stops them. Dev servers and other background shells an agent starts keep running after the turn ends and when the chat's workspace folders or worktree change, so they still answer in the next turn. Archiving the chat stops them instead of leaving them running.
  • A cleaner, more compact agents sidebar. Agent rows in the agents sidebar now take up a single line, so more agents fit at a glance, and they can show when each agent was last updated. A new Search row opens the chat history picker in place of the inline search box, and keyboard shortcuts for rows appear only when you hover.

Privacy and security

  • Admin Subagents respect your Private Inference model choices. With Private Inference, subagents no longer run on models you turned off in Settings > Models. Custom subagents pinned to a turned-off model run on the parent agent's model instead.
  • Admin Security hardening for git diffs, links, and repository blocklists. Viewing diffs, blame, or history no longer runs a repository's configured external diff or textconv command, so opening an untrusted repo can't execute code. Wildcard trusted domains in link protection no longer match lookalike hosts, and agent shell commands that list directories now respect admin repository blocklists, including ** globs.

Fixes

  • Accessibility fixes. Images in chat can be reached with Tab and opened with Enter or Space, and screen readers now name the agent status icon, the plan Build options menu, and the built-in browser by its page title. In Cursor Settings, the Partial Accepts tooltip shows the actual keybinding, the Open and Invite buttons have names, and Tab reaches a model's switch before Remove.
  • Smoother Cloud Desktop viewer. The Cloud Desktop viewer for cloud agents decodes screen updates more efficiently, using less CPU.
  • Fixes to agent chat scrolling and rendering. Long agent chats no longer jump while you scroll, and loading a large cloud agent conversation no longer blocks the editor. Streamed messages stay in one bubble, user messages no longer show internal automation tags, and code blocks, lists, and equations no longer add extra gaps between messages.
  • Fixes to agent tools. Agent sessions no longer end when the model sends an out-of-range value to Shell, Grep, or Read, and long todo lists no longer crash the agent. Glob now finds empty and binary files, edits no longer change only the first of several overlapping matches, and web search and mode switch calls show up in chat before they ask for approval.
  • Fixes to the agent chat input menus. The / menu no longer lags when many commands match, and typing after choosing /add-plugin no longer freezes the input. Arrow keys now wrap around in the @ and + menus, closing the + menu returns focus to the + button, and tiny images in the image preview are enlarged so they're readable.
  • Fixes to the @ mention menu in chat. Late-arriving results no longer move the row you've highlighted, rows no longer flash while results load, and slow sources still show up. For cloud agents, including when editing a sent message, the menu shows the last known files right away, refreshes once the file list loads, and only suggests folders the agent's machine has.
  • Fixes to Review Changes. Opening a large GitHub pull request in Review Changes no longer freezes the editor. Review Changes also uses less memory and stays responsive on long agent chats that touched many or large files.
  • Fixes to loading local media files. Local video and audio files now play and seek correctly in the editor. AVIF, ICO, WebM, and M4V files from anywhere on disk now load in previews, and files under a symlinked home or extensions folder are no longer refused.
  • Fixes to saving Markdown files from the preview. Saving a Markdown file from the Markdown preview editor without editing it no longer rewrites its formatting. Edits made in the preview no longer remove the file's final newline or add a blank line after a trailing list, code block, table, or heading.
  • Fixes to model settings. The Explore Subagent Model help in Settings no longer shows two tooltips at once, and the model picker no longer shifts sideways when you change the selection.

Agents

  • Teach your coding agent to use the Origin CLI. origin agent prompt prints a Markdown guide to origin for AI agents, matched to your installed version, with focused topics such as pull requests, review, and CI. origin agent setup installs a skill for Cursor, Claude Code, or Codex that points the agent to that guide, and mistyped commands suggest the closest match. See agent commands.

Pull requests

  • origin pr merge stops merges that would land on a closed pull request's branch. If a pull request targets the head branch of a pull request that was closed without merging, origin pr merge refuses instead of landing your change on that branch. The message tells you how to fix it: retarget to trunk, reopen the parent and stack on it, or pass --base-ok to merge anyway.
  • Merge-conflict status no longer shows a pull request as clean too early. origin pr view --checks, origin pr watch, and origin pr status -c now say when Origin is still checking for conflicts, is checking a conflicting pull request again, needs a restack, or has nothing to merge. Before, all of these showed as no or MERGEABLE.
  • Merge blockers say what to do next. Blockers in origin pr requirements now give the next step, such as "Restack the stack from #N to continue." or which branch to retarget onto when a stack's base branch was deleted. A blocker on a lower pull request in the stack names that pull request.
  • Shorter Gates section in origin pr requirements. It shows only the gates your pull request requires, with their status at head, plus any errors in the head's .origin/gates.ci. When nothing is required, it prints Gates: none required.
  • origin pr search reports every invalid qualifier at once. When a search has several unknown or malformed qualifiers, the error lists all of them instead of making you fix them one at a time.

Stacks

  • Checking out a stack updates its trunk and protects teammates' branches. origin pr checkout --stack and origin stack get fetch the stack's trunk branch and fast-forward it when it is behind. Branches for pull requests you didn't author are frozen, so stack commands won't rewrite them; run origin stack unfreeze <branch> to take one over. See track and freeze.

Repositories

  • Imported rulesets keep merge methods. origin ruleset import --from-github carries a GitHub ruleset's allowed merge and squash methods onto the imported pull request rule. Rebase is dropped with a warning. See rulesets.
  • Faster default-branch lookup in origin repo clone-fast. It asks the server for the repository's default branch alone instead of downloading the full list of refs, and falls back to the old lookup if that request fails.

Reliability

  • Git stops with a login hint instead of asking for a username. When you aren't logged in to Origin, or your credentials expired and can't be refreshed, git operations through the Origin credential helper stop with a message to run origin auth login. Before, git fell back to a username prompt.
  • Fewer failed reads right after you create a pull request. When Origin briefly reports that it is unavailable and asks the client to retry, origin commands that only read data try again, up to four attempts, instead of failing. That covers the moments after you create a pull request, before its first version is readable. Commands that change data are never retried.
  • Usage errors show what you can run next. A mistyped command or missing argument prints the error with a short usage line, the commands and flags available at that level, and a pointer to --help. -h now works as --help, and origin config get-channel names the environment variable that set your channel.

Fixes

  • Fixes to stack restack, checkout, and get. origin stack get no longer fails on every run claiming --take-remote and --keep-local were both passed. origin stack restack rejects a branch with a restack <branch> first message instead of dropping commits. Stack checkout follows branches you renamed locally and reports when the stack's base branch is gone.
  • Fixes to origin pr watch. It no longer reports a failure when the server hasn't read the pull request's merge rules yet; it waits up to two minutes instead. It also stops reprinting status that hasn't changed.

Web

  • Admin Restrict merge methods per branch with rulesets. In a repository's Settings > Rules and Protections, the ruleset dialog has a new Restrict merge methods option that limits pull requests on matching branches to merge commits, squash merging, or both. The merge button only offers methods the target branch allows. When no method can land, the pull request shows a blocker that names the branch, the rulesets involved, and how to fix it, instead of failing at merge time.
  • Choose how auto-merge lands your pull request. When you enable auto-merge on a pull request, you can now pick squash or merge commit, and the commit title and message you edited are used when it lands. The button shows your pick, for example "Enable auto-merge (squash)", and once auto-merge is on, the merge popover shows the chosen method and commit.
  • Admin Block large files with a maximum file size rule. In a repository's Settings > Rules and Protections, you can now create a Repository ruleset with a Maximum file size limit of up to 100 MB. Origin rejects any push that adds a file over the limit and names the files that are too large. The ruleset list summarizes each rule by its limit, for example "Files up to 5 MB". The rule is also available through the Origin API and Go SDK; see the Origin API changelog for details.
  • Admin Require approvals from specific groups before merging. In a repository's Settings > Rules and Protections, a ruleset can now require approvals from specific groups, each with its own minimum number of approving members, chosen from an Add group menu. A pull request blocked on this shows a blocker naming the group, with how many approvals it still needs and how many it has, and suggests requesting review from that group. Group rows show the group's name, type, and member count, and groups that can't be named read "Deleted group" or "Group outside this organization."
  • Admin Choose how code owners review pull requests in rulesets. In Settings > Rules and Protections, the ruleset editor now has a Code owner review setting with Off, Classic, and Agentic options. It replaces the "Require review from code owners" toggle. Classic requires an approval from a code owner for the files they own. Agentic still enforces CODEOWNERS, but when Origin's agent reviews ownership on a pull request, its decision replaces the approval count. When a change is blocked, Origin shows which owners need to approve and for which files.
  • Admin Apps can be code owners. You can name an app as a code owner in your ORIGIN_CODEOWNERS file with app_<id>, and an approval from that app satisfies the paths it owns, provided the app is installed on the repository's namespace. On the pull request page, app owners show their name and avatar. Apps that aren't installed appear muted with a "Not installed" hint, and the codeowners preview explains why they can't approve.
  • Admin Ruleset saves no longer overwrite someone else's changes. In Settings > Rules and Protections, if someone changed a ruleset after you opened it, saving now asks you to reload first. Descriptions can span multiple lines, and pushes blocked by a ruleset name that ruleset. Active rulesets are now always enforced, even in repositories with many disabled ones.
  • Admin See which rules were skipped when a pull request merges with a bypass. When someone on a ruleset's bypass list merges a pull request without meeting all its rules, the pull request's activity now shows an entry like "merged, skipping 2 rules" that lists each skipped rule, such as required reviews or required status checks. The bypass is also recorded in audit events.
  • Large pull requests load every file. The Files changed view now keeps loading files on large pull requests instead of stopping at 500, including when you compare two versions. If a list still stops short, a line tells you how many files are shown.
  • Exclude, date-filter, and label-filter pull requests in the list search. On a repository's pull requests list, you can now exclude results with negated qualifiers like -author:@me, -label:bug, or -is:draft, and filter by date with created:, updated:, merged:, and closed: using dates, comparisons, or ranges such as merged:2026-09-01..2026-09-30. Rows now show their labels, and clicking one adds it to your search; label:a,b matches either label. Merged and closed rows also say when they merged or closed, and loading more pages in large repositories is faster.
  • Keyboard shortcut sheet on pull requests. Press ? on a pull request page, or choose Keyboard shortcuts from the overflow menu, to see every available shortcut, including the keys for switching tabs and moving through the diff.
  • Discard a pending review. When you have pending review comments on a pull request, the review menu in the pull request header now offers Discard review. After you confirm, Origin deletes your pending review and all its draft comments.
  • Symbol navigation and a Raw button in codebase file views. When you browse a file on a wide screen, hover a symbol to see its signature, or click it to jump to its definition and references. File views also have a Raw button, next to copy and download, that opens the file's plain contents in a new tab.
  • More reliable code tours. Code tours that fail to generate now retry once automatically, and if a default tour fails partway, you see the saved tour instead of an error. In a whole-stack tour, each chapter shows the change made by that chapter's pull request, and a Files you own tour follows the files you own when you switch comparisons. When a Cloud Agent edits a pull request's tour, you see that edit instead of an older tour you generated.
  • Jump to a pull request author's Cursor profile. On pull request pages, author names and avatars in the timeline, comment threads, and hover cards now link to the person's public Cursor profile when they have one. Authors in a pull request stack now appear the same way they do on the pull request itself, and profile links no longer disappear while the page updates.
  • Admin Ruleset bypass for apps pushing directly is clearer. An app on a ruleset's Bypass list can now push directly to a branch that blocks direct pushes when it uses its own installation token. An app pushing with a person's token no longer gets that person's bypass, so it has to open a pull request instead.
  • More reliable Git LFS downloads, with clearer errors when a file is missing. In repositories mirrored from GitHub, Git LFS downloads now fall back to GitHub when a file isn't in Origin yet. When a file is missing everywhere, the error names the object and explains how to upload it with git lfs push or skip it with GIT_LFS_SKIP_SMUDGE=1. If GitHub rate-limits a download, Git LFS is told to retry later.
  • Push-triggered automations now know who pushed. Automations triggered by a push to an Origin repository now get the pusher as the actor, the same as pull request, comment, and label events already did. Actor-based filters now match valid pushes instead of skipping them. If the pusher is hidden or can't be resolved, the automation still won't run.

API

  • Manage auto-merge from the Origin API. You can now turn auto-merge on or off for a pull request through the Origin API, see who turned it on and how it will merge, and get webhook events when auto-merge is turned on or off. These endpoints and events are in preview. See the Origin API changelog for details.
  • Create a repository through the Origin MCP server without picking an owner. The create_repository MCP tool now works without an owner: Origin sets up your personal namespace the first time and creates the repository there. This needs an eligible plan and a Privacy Mode setting that allows code storage. Team members and app tokens still need to pass an owner.
  • New webhook events for drafts, authorship changes, and repository labels. Origin apps can subscribe to pull_request.unpublished, sent when a pull request is moved back to draft, and pull_request.commandeered, sent when someone takes over a pull request's authorship, with the previous author in the payload. Apps with read access to repository labels can also subscribe to repository.label.created, repository.label.updated, and repository.label.deleted. Pausing or revoking an app now stops installation webhook deliveries right away, and reviewer webhooks are no longer sent with an empty reviewer. See the Origin API changelog for payload details.
  • Create missing labels while labeling a pull request through the Origin API. Adding or setting pull request labels through the Origin API and Origin SDK now accepts create_missing, so automations can apply a label that doesn't exist yet in the repository without a separate create call. Existing labels match by name regardless of case, and creating labels requires permission to manage labels. Without the option, the error now lists every missing label name. See the Origin API changelog for details.
  • Origin MCP tools accept GitHub-style arguments. Agents and scripts written for GitHub's MCP server now run against the Origin MCP server with fewer errors. Tools accept owner/name strings, full_name, nested repository objects, and ref in place of sha. The tree, commit, and pull request read tools also work without a sha or view, defaulting to HEAD, metadata, and the summary.

Fixes

  • Generated-file markers in diffs respect linguist-generated=false. Files marked linguist-generated=false in .gitattributes are no longer shown as generated in pull request diffs.
  • Fixes to Git pushes and repository reads. Pushes evaluate rulesets with many wildcards quickly, a recreated repository no longer serves the old repository's refs, and commits with large messages now return their metadata.
  • Fixes to team and namespace setup. Merged teams that have members but no seats are no longer treated as inactive. Auto-merges no longer stall, review reminders and Slack DMs arrive, and the team's privacy mode applies. More plans can now write to their personal namespace without an upgrade prompt, and automatic namespace setup picks an available name when your preferred one is taken.
  • Pull request pages no longer show checks from a different base. Checks reported against a different base no longer appear on a pull request page, including as live updates arrive or after the base moves.

Projects

  • Project coordinators can start workers wherever your code lives. A Project coordinator can list your environments and repositories, then start a worker in the environment that holds the code, or directly on a repo using its default setup, even when the Project has no repository of its own. Environments and repos outside the Project need your approval from a card that names them, and approved ones appear in the Project's allowed list, where you can revoke them.
  • Admin Project agents ask before using a new machine or environment. For every team, a Project's agents now ask a person to approve the first time they use a Self-Hosted Machine, pool, or environment. Team admins can turn on Automatically allow Projects to place their work in the Dashboard under Cloud Agents → Security. Approvals you already granted still apply, and an approval for a repository also covers its GitHub or Origin mirror.

Agents

  • Messages sent as an agent finishes are answered in the same run. If you send a message to a Cloud Agent just as it is wrapping up, the agent picks it up as one more step instead of queuing it as a separate follow-up.
  • Start Cloud Agents from a short commit SHA. You can start a Cloud Agent on a GitHub repository from an abbreviated commit SHA of 7 or more characters, like the ones git log --oneline prints, instead of getting a "branch or commit not found" error. Branch and tag names with the same text still take precedence.
  • Agents can update pull request descriptions they didn't write. When you ask a Cloud Agent to update a GitHub or Cursor-hosted pull request it didn't open, it now adds its content below the existing description instead of refusing, and it can fill in an empty description. Agents in a single-repository environment no longer need the repository named to create or manage pull requests, and completed pull request actions show as finished in the conversation.
  • A clearer agents sidebar on pull request pages. The agent list in the floating dock on cursor.com pull request pages opens right above the input and shows how many agents are working or need attention. When the agent you have open is idle but agents under it are still working, its header says "Waiting on" those agents. The sidebar shows a Retry notice instead of loading forever, or tells you when Privacy Mode (Legacy) blocks Cloud Agents.
  • Admin Teammate follow-ups on agents without a repository follow your team settings. Teammates can follow up on a colleague's Cloud Agent that runs on a personal environment with no repository, as your team's Team Followups and Allow VM Sharing settings allow. Before, those agents were read-only to teammates. A teammate who opens or follows up on someone else's agent must also have access to every repository the agent was granted.
  • Admin Clearer messages when usage limits block a Cloud Agent. When a usage limit stops a Cloud Agent from starting or taking a follow-up, the error tells you what to do next: turn on on-demand usage, raise your own or your team's on-demand limit, or, on a trial, switch to a Cursor model or Auto. If you switch a follow-up to Auto after a trial limit, the agent continues instead of being blocked again.
  • Admin Clearer answers when a team model block turns off computer use. If your team admin has disabled a model that computer use depends on, the Cloud Agent tells you that computer use is unavailable, which model is blocked and what it's used for, and that an admin can re-enable it in team model settings.
  • Open Settings with a keyboard shortcut. Press Cmd+, (Ctrl+, on Windows and Linux) on the agents pages to open Settings. The collapsed sidebar no longer pops open when you hover over it; click the Open sidebar button instead. Opening a conversation only collapses the sidebar on windows narrower than 1024px.

Environments

  • Rename environments and find config-as-code ones on the dashboard. Rename an environment from its detail page with the pencil button next to its name. Environments defined by a committed environment file show a Config as code badge, and a Config as code switch filters the list to them. When a build fails because the GitHub App can't access a repository, the banner names the affected repositories and links to your GitHub installation settings.
  • Environment builds no longer silently use the wrong environment.json. When an environment's tracked environment.json is missing, empty, or can't be read, the build now fails with an error that names the file and repository, instead of quietly falling back to .cursor/environment.json or an older saved config. New environments track the repository's environment.json only when you choose that option.
  • Fewer duplicate and overwritten environments. Creating an environment from the dashboard or promoting a personal environment to your team now always creates a new environment instead of overwriting one with the same name or repositories. If a team environment with that name already exists, you're asked to rename yours first. Agents started on the same set of repositories reuse one environment.
  • Admin Environment build cache is encrypted with your customer-managed key. For Enterprise teams with a customer-managed key, the Cloud Agents environment build cache is encrypted with a key your customer-managed key controls. If you revoke access to your key, or you have no key in the build's region, environments build without a cache.

Self-Hosted Machines

  • Agents on Self-Hosted Machines handle offline machines better. When a pool machine running your agent goes offline, the agent waits for that same machine to come back and continues there, and moves to another machine only after that. If you start an agent or send a follow-up to My Machine while it's disconnected, you get a machine-offline message after about 15 seconds instead of waiting for minutes.

Integrations

  • Jira automation rules and GitHub Enterprise @cursor comments launch Cloud Agents correctly. Jira automation rules start a Cloud Agent with the rule's prompt as its task, including rules with no work item attached. @cursor comments on GitHub Enterprise Server repos now work, and run under the team that owns the GitHub app when you belong to it.
  • Azure DevOps repos work as org/project/repo shorthands. You can point a Cloud Agent at an Azure DevOps repository with a three-part shorthand like my-org/My Project/portal, for example when switching repos or setting a default repo in Microsoft Teams, instead of pasting the full dev.azure.com URL.
  • Admin Cloud Agent commits in OpenTelemetry Export. Enterprise teams using OpenTelemetry Export receive a cursor.cloud_agent.commit log for each commit a Cloud Agent pushes to a GitHub repository. Each log includes the commit SHA, the repository URL, the number of files changed, and line counts when available, and links back to the agent run.

MCP

  • Admin Cloud Agents follow your team's plugin and MCP allowlist. When an admin turns on Team list enforcement, Cloud Agents connect only to plugins and MCP servers on the Team list, and the tool restrictions set for each server apply. When enforcement is off, those tool restrictions no longer block other MCP servers.
  • Cloud Agents no longer get stuck on sign-in cards for command-based MCP servers. Cloud Agents no longer pause on an MCP sign-in card for stdio servers, which run as a command on the agent's machine and have no Cursor sign-in. Before, the run waited until you clicked Skip or the card timed out. Cloud Agents also no longer ask you to sign in to a server that is still loading.

Fixes

  • Fixes to agent runs. Stopping an agent while it continues after a stop hook now works, parallel file writes in one step no longer overwrite each other, and agents no longer get duplicate AGENTS.md or CLAUDE.md instructions. Follow-ups sent during a CI-fix turn run as you, and Bugbot's Fix in Web continues the existing pull request instead of opening a new one.
  • Fixes to environment builds and network allowlists. Environment setups longer than an hour no longer fail partway for teams using PrivateLink. A Cloud Agent whose VM is recreated on a follow-up boots the environment's latest successful build, and changing a team environment's egress settings triggers a rebuild. One malformed network allowlist entry is now skipped instead of breaking or widening the whole policy.
  • Fixes to repository access. Cloud Agents on GitLab instances that require Git protocol v2 now start, brief GitLab outages are retried, and failed access checks show GitLab's actual response. Tokens or passwords embedded in repository URLs are stripped before they're stored.
  • Fixes to Project coordinators and workers. Workers that a coordinator starts with a variant of its own model keep the coordinator's context window instead of jumping to 1M. Coordinators with more than 1,000 workers list their newest workers, and renaming a Project before its first turn now sticks.
  • Fixes to Self-Hosted Machines. Runs that land on a machine that is shutting down start on another machine in the pool, and pool machines left behind by a deleted run reset so the next agent can claim them. Agents recover from a lost machine connection by reconnecting on the next step, and Windows machines resolve skill, plugin, and workspace paths correctly.
  • Fixes to MCP connections. MCP servers that don't need sign-in no longer have their tools disappear between listings, and servers whose sign-in was revoked or expired show as needing authentication instead of exposing cached tools. A newly added server or tool is no longer reported as missing, and an expired server session reconnects instead of failing the call.
  • Fixes to code tours. Code tours for pull requests in tall stacks now generate instead of failing, and a tour that fails mid-generation ends with an error instead of leaving the viewer waiting.
  • Fixes to recording annotations. Large screen recordings now get chapters, a thumbnail, and trim points instead of being skipped, and annotations are no longer lost when another save happens at the same time.
  • Fixes to the web app. The agents navigation drawer no longer opens empty at fractional browser widths, and you see all of your pinned agents, not just the first 1,000.
  • Fixes to Slack replies. Agent replies in Slack keep dollar sequences like $1, $&, or $$ exactly as written instead of repeating or dropping parts of the message.

Integrations

GitHub

  • Admin Bugbot reviews pull requests in shared GitHub installations. When another team in your organization shares its GitHub installation with your team, Bugbot runs on pull requests in the shared repositories using your team's Bugbot settings.
  • Admin Connecting GitHub no longer turns on Bugbot for every repository on Enterprise teams. When an Enterprise team or a GitHub Enterprise Server installation connects GitHub, existing repositories stay off until an admin turns Bugbot on for them. New repositories also stay off unless you change the installation's auto-enable setting, and these changes appear in your team's audit log.
  • Admin Clearer GitHub access errors. When an agent can't reach a GitHub repository, the error says why. A repository not granted to the Cursor GitHub App comes with a link to grant access, an organization that requires SAML SSO comes with a link to authorize, and IP allow list blocks are explained. Changing settings for an installation linked to a different Cursor account explains how to uninstall and reinstall it from your team's dashboard.
  • Admin Fixes to the GitHub integration. Admins of merged teams can manage GitHub and Bugbot installations again instead of seeing "Team is not active". Agents with access to more than 500 repositories get a GitHub token instead of failing, and GitHub requests no longer fail when HTTP/2 isn't available. When GitHub rate limits token refreshes, you see a "try again later" message.

Bitbucket

  • Admin Bitbucket Data Center explains a missing permission. If Cursor's application link grants less than Repositories Read, connecting Bitbucket Data Center shows a page asking your Bitbucket admin to raise it before you reconnect, and the setup instructions list that permission.
  • Admin Fixes to Bugbot for Bitbucket. Bitbucket Cloud admins who get admin access through a group or workspace role can turn Bugbot on or off for a repository. Reconnecting a repository no longer fails with 405 errors when Bugbot updates its existing webhook.
  • Fixes to Bitbucket. Repository lists in large Bitbucket workspaces load partially instead of failing, and reading pull request diffs and repositories no longer fails when Bitbucket Cloud redirects. Creating a pull request reuses the open one even when the repository name differs in letter case, and empty secondary repositories are skipped.

GitLab

  • Admin GitLab credentials stay on your GitLab host. When a GitLab instance redirects a request to a different host, Cursor refuses the redirect with an error instead of sending your token or OAuth client secret there. Same-host redirects, such as a moved project or renamed namespace, are still followed.
  • Fixes to the GitLab integration. Empty GitLab repositories added as secondary repositories are skipped instead of making the agent fail. Webhooks with a missing token or unknown registration return an error, so GitLab shows the failure instead of reporting success.

Dashboard and admin

Plugins and MCP

  • Admin See which servers Team list enforcement would block before you turn it on. Team list settings on the Plugins page, and the new Turn On Team List Enforcement dialog, now list the plugins and MCP servers members use that aren't on the Team list, with how many members use each. You can add a plugin to the Team list or allow a remote MCP server in one click, and the dialog warns you when enforcement would block every MCP server. See MCP server trust management.

Integrations

SSO and SCIM

  • Admin Control who can join through SSO. Organization admins can turn off SSO auto-enrollment in organization settings. When it is off, only invited or SCIM-provisioned users can join, while existing members keep signing in with SSO as before. Anyone else who signs in through your SSO connection is told to ask an admin for an invite. See SSO settings.

Members and groups

  • Admin Create team groups that the whole team can see. The New group dialog on the Members Groups tab now has a Visibility switch, so admins can make a group and its members visible to the whole team when they create it instead of editing it afterward. Groups stay visible only to admins unless you turn the switch on.
  • Admin Choose how many members to show per page. The Members page now has a Rows per page control at the bottom of the list, so admins of large teams can show 50, 100, or 250 members at a time. Your choice is kept in the page URL, and changing it takes you back to the first page.

Agent settings

  • Admin Project agents now ask before using a new machine, on every plan. The Automatically allow Projects to place their work switch in Cloud Agents security settings now starts off for every team, not only Enterprise teams, so a person approves each new machine, environment, or repository a Project's agents use. Teams that turned the switch on keep auto-allow, and approvals already granted stay in place. See Cloud Agents security settings.
  • Admin Clearer warnings for network allowlist entries that won't match. In the team Network access settings, admins now see which sandbox allowlist entries cloud agent egress ignores, along with explanations for unsupported characters and wildcard patterns. Catch-all entries such as *. and URL paths are no longer accepted for cloud agent egress, so a mistyped entry can't open access to every host. See network access settings.
  • Admin Choose a mode for Approval Agents. Approval Agent settings now include a Mode option under Configuration. Default is the faster choice and fits most teams. Low costs less and runs slower. You need edit access to change it. See PR Routing & Approval.

Rollouts

  • Admin Send Rollouts notices to Slack channels by repository, service, or environment. In Rollouts settings, team admins can subscribe Slack channels to change notices that match a repository, service, or environment, and add an optional message, including @mentions, that's posted in the thread after each new issue. Members can add their own personal subscriptions, and both team and personal subscriptions can be edited later. See Rollouts notifications.
  • Admin Send deployment events to Rollouts from your GitHub deployments. When you set up deployment events in Rollouts, you can choose Use my GitHub deployments to record deployments from your watched repositories, with no CI reporter needed. The manual setup now shows numbered, collapsible steps, and its copyable snippets report a missing or rejected Cursor API key in your CI log without failing the deploy. Deployments from gitlab.com projects now appear on a deployment's page, with Version and Previous version linking to the GitLab commit. See sending deployment events.

Cloud agent environments

  • Find, filter, and rename Cloud Agent environments. The Environments list now marks environments defined by a committed .cursor/environment.json with a Config as code badge, and a Config as code switch shows only those environments. You can rename an environment by clicking its name on the environment's edit page, and the Repo filter has a search box. In the default environment picker, unnamed environments show their repositories, such as "web-app +3", instead of an ID, and you can search for them by repository name. See environment management.
  • Clearer messages when a Cloud Agent environment build can't use its environment.json. When an environment build fails because its tracked environment.json is missing, empty, or can't be read, the build page now says so and tells you to restore or check the file, then rebuild. Agents keep using the last successful build in the meantime. Builds that fail because the GitHub App can't access a repository now list the affected repositories with an Open installation settings button, and teams that require self-hosted workers see a note in place of the Trigger build action. See Cloud Agent Builds.

Usage and analytics

  • Admin Export usage events for your whole organization. Organizations with pooled usage now get a Usage events section on the organization Usage page, with an Export button that downloads event-level usage across the organization's teams for the selected date range and filters. If an export isn't possible, for example when the organization has both merged and unmerged teams, the button is disabled and the page explains why.
  • Reorder usage event columns and read usage labels more easily. On the Usage page, you can drag the usage events table's column headers, or move them with the keyboard, into the order you want. Your browser remembers the order, and Date and User stay in place. Usage types now read "Additional usage" and "Credit grants", Auto requests show as Auto or as the routed model with its Auto option in the chart and model filter, and usage owned by a team rather than a person or API key shows as Team in the table and CSV export.

Security and audit logs

  • Admin More team changes in the audit log. The audit log now records changes to cloud agent network egress and sandbox networking settings, and to your team's repository allowlist and blocklist, including repo patterns. It also records Remote Control events, such as a computer being registered or a device being paired, unpaired, or revoked, with a new Remote Control filter. Pull request takeovers are now logged too, and takeovers show the previous author. See audit logs.

Workspace settings

  • Admin Add an icon to your team or organization. Admins can upload, change, or remove a team icon (PNG, JPEG, or WebP up to 2MB) in the Team profile settings. The icon appears in the workspace switcher, command palette, team selector, and the organization's teams table. Organization owners can also set an organization icon from the organization Overview page.

Automations

  • Admin See more flagged vulnerabilities per page. The flagged vulnerabilities table on the Vulnerability Scanner page now has a rows-per-page selector with 10, 25, 50, or 100 rows. It shows 25 by default instead of 5. Long titles wrap to two lines, and the Prev and Next buttons no longer move when you change pages. See Vulnerability Remediation.

Account and navigation

  • See what you lose before deleting your account. The delete account dialog in Profile settings now lists what will be deleted, starting with your Cursor account's name, email, and member-since date. If your account is merged with a Grok account, it also shows that account under Deleted and your linked X account under Unlinked, since X accounts are unlinked, not deleted. The dialog also notes that payments already made are not eligible for a refund.
  • Reset a saved color palette from Appearance settings. If you saved a colorblind or high-contrast color palette, Settings > Appearance now shows it in a Color palette row with a Reset button. Resetting shows a toast with Undo, so you can restore the palette.

Fixes

  • Admin Fixes to Rollouts setup and issue pages. Rollouts setup now pre-selects your team's existing observability connections, and Rollouts only shows as Enabled once setup is finished. If Rollouts is already on, setup still shows that step, with a link to its settings. Clicking Fix on an issue opened from a direct link now starts from the right rollout, and links to pull requests open on the portal you're using.
  • Fixes to cloud agent environments. Restoring an older team environment version no longer removes your personal override, and the Secrets row now says it includes team and personal secrets, with Manage opening the secrets that match the environment's scope. If the environments list fails to load, the Environments pages, Slack repo routing rules, and the agent environment picker now show the error or fall back instead of loading forever. The Environments section also recognizes any connected source control provider, so it loads faster and no longer shows the Connect Source Control card by mistake.
  • Admin Fixes to usage and analytics. Team usage event tables now load long date ranges, such as 30 days, without timing out on large teams. The usage card shows a member's own spending limit, including a $0 limit, instead of the team limit. The Usage Leaderboard no longer lists service accounts, team API keys, or unauthenticated usage, and no longer merges members who share an email or have none.
  • Admin Fixes to team invites and team selection. Copy invite link now uses the role you pick in the invite form on teams without seat tiers, so an admin link no longer adds people as members. When you belong to several teams, paid teams now come before free ones, so you no longer land on a free team by default.
  • Fixes to integrations and shared pages. Teammates who haven't connected Slack now see a deployed agent's existing Slack apps as read-only rows with a Connect Slack button, not an empty state. Shared canvases now print on white paper without the browser's header and footer, and the disabled Team sharing option explains when your team's admin turned it off.
  • Fixes to dashboard navigation and phone layouts. On phones, tab strips scroll to the screen edges, tap targets are larger, and the members search and setup checklist no longer crowd the layout. The Cmd+K command palette now opens above other dialogs, and closing it leaves the dialog underneath open. Each dashboard tab sets its own browser title, a failed background refresh no longer replaces a loaded workspace with an error screen, and profile settings show your saved first and last name.

Automations

Triggers

  • Label triggers for Origin pull requests. Automations on Origin repositories can now use the Label change trigger, which runs when a label is added to or removed from a pull request, as it does for GitHub.
  • Inline review comments on Origin pull requests now trigger automations. Automations with a pull request review comment trigger now run when someone leaves an inline review comment on an Origin pull request. Before, those comments only reached automations that trigger on top-level pull request comments.

Security agents

  • Admin Vulnerability Scanner caps hourly scans per team. Vulnerability Scanner automations run at most 500 repository scans per team per hour. Scans over the cap are skipped and show in run history as rate limited.
  • Subagents in managed security runs no longer message on their own. Fixed an issue where helper subagents in managed security review, vulnerability scanner, and PR approval runs could send Slack and Teams messages or resolve review threads on their own. Those actions are now refused for subagents in these runs.

Integrations

  • Automations load the MCP servers you named. Automations that pick an MCP server by name, like slack or google-drive, now find it even when the capitalization or spacing doesn't match. Before, these runs could start without the server's tools. If Cursor still can't find a server, the run shows a warning telling you to check the name or reconnect the server under Integrations.
  • Clearer status when GitLab has MCP turned off. When a GitLab MCP server in an MCP action is blocked because GitLab has MCP turned off for your groups, the automation editor now explains that an Owner of your top-level GitLab group must turn on MCP client access, instead of showing a sign-in prompt that can't succeed.
  • Admin Only team admins can connect a team's PagerDuty integration. Connecting, replacing, or moving a team's PagerDuty integration now requires a team admin, the same as disconnecting it. A PagerDuty account connected to one team can't move to another team until an admin disconnects it. The connection page now tells you why a connection failed: your session ended, you aren't a team admin, or the account belongs to another team.

Automation runs

  • Admin Managed automations no longer run on models your team blocks. When your team's model settings block every model a managed automation can use, such as the security reviewer or vulnerability scanner, the run now fails with a "Model blocked" error asking a team admin to allow a model. Before, the automation could fall back to a blocked model.

Automations page

  • The Model column shows the variant your automation actually runs. The Automations list, overview table, and run history now label each automation's model with the effort and Fast settings you saved, instead of the model's default variant. If a saved selection doesn't match a single variant, the column shows the plain model name.

Fixes

  • Fixes to which branch and repository automations launch on. Automations tied to a single-repo environment, including repos saved as SSH URLs, now run on the branch you saved instead of the environment's default branch. Automations that target a specific self-hosted machine fall back to main instead of failing when Cursor can't reach the git host. Runs that fail because the GitHub App was installed under another account now say the repository isn't connected to this team and explain how to reinstall the app.
  • Fixes to the automation editor and integrations. MCP actions now show when a server needs sign-in even if no sign-in link came back, and clicking sign in gets a fresh link. PagerDuty services load faster on accounts with many services, and descriptions over 1,024 characters now show a clear error. Turning off Change Monitor reports a failure instead of looking successful while it keeps running, and the Bugbot MCP tools and installation pages under From Cursor now load.
  • Automations list rows work with the keyboard. You can now reach and open rows in the Automations list with the keyboard. Tab moves focus from row to row.
  • Fixes to long automation runs. Runs that take a very large number of actions no longer fail to save their action history. The oldest entries are compacted so the run keeps recording.

APIs

  • Admin Make directory groups visible to your whole team. Team directory groups in the Admin API now have a visibility field. Set it to team when you create or update a group so every team member can see the group and its members, or team_admin to keep it visible only to admins. If you leave it out on create, the group is visible only to admins.
  • Admin Fixed missing member spend in the team spend endpoint. POST /teams/spend now counts each member's billed usage in full for teams on combined billing, so reported spend no longer comes up short. If spend data can't be read, the endpoint returns a 503 error instead of reporting zero spend.

Security and approval agents

Security agents

  • Vulnerability Scanner findings show in Vulnerability Remediation. Fixed an issue where findings from a Vulnerability Scanner didn't appear in the Vulnerability Remediation section.
  • Show more flagged vulnerabilities per page. The Flagged Vulnerabilities list now has a Rows per page selector (10, 25, 50, or 100) and shows 25 rows by default instead of 5. Long titles wrap to two lines, and the Prev and Next buttons stay in place as you page.
  • Fixed multi-file findings close when you push a fix. When a finding names more than one file, Security Reviewer re-checks each of those files on every push, so a fix in any of them closes the finding. You no longer need to resolve it by hand. Open findings in files you haven't changed now count toward the security check, so the check status matches the open threads.
  • Saving large Security Reviewer configurations works. Saving a Security Reviewer with many selected repositories, exclusions, or Slack channels no longer fails.

Approval agents

  • Approval comments show everyone who can approve. When a PR Routing & Approval policy is waiting on review and more owners can approve than were requested, its comment reads like "Needs 1 approval from any of: @alice, @bob (+2 more)". On Origin repositories, a pending policy that lists bot owners names them and shows how to request their review from the Reviewers list or with origin pr reviewer request.

Models and Cursor Router

  • New model: Gemini 4 Argon. Gemini 4 Argon is now available in Cursor.

Teams and installations

  • Admin Bugbot runs under the right team on shared GitHub installations. When a GitHub installation is shared across Cursor teams, Bugbot reviews on a repository run under the team it was granted to, using that team's settings and plan, instead of the installation owner's or the PR author's. The Bugbot repository list shows an "Enabled by" badge naming the team on repositories another team already runs Bugbot on, and explains why you can't enable them there.
  • Admin Manage Bugbot on GitHub installations shared with your team. If another team in your organization shares its GitHub installation with your team, you can change per-repository Bugbot settings from the dashboard without taking over the installation. The installation details page marks these installations as Shared and names the owning team, which still manages installation-level settings.

Reviews

  • Local reviews work in Mercurial repositories. When you run Bugbot from your agent or run a security review on local changes, it now reviews Mercurial checkouts as well as git repositories, without running the repository's hooks. If a folder has neither, the review says so and names the path. If your Mercurial client is too old to skip hooks safely, it asks you to upgrade.
  • Rule violations are no longer dropped when a pull request calls them intentional. Bugbot now keeps a finding that correctly flags a violation of one of your team, project, or manual rules, even if the pull request description says the change was deliberate. Findings that cite a rule the diff doesn't break are still filtered out.
  • Long local reviews finish with results. When you run Bugbot from your agent and a review runs unusually long, Bugbot stops investigating at a set limit and replies with the bugs it has already verified, instead of running on or ending without an answer.

Settings and access

  • Admin Restrict Access works with pull requests from Cursor agents. When a Cursor agent opens a pull request, the team's Restrict Access allowlist or blocklist now checks the GitHub account of the person who launched the agent, not the bot. When Restrict Access blocks a review, Bugbot's "couldn't run" comment says which list blocked it and how an admin can fix it, instead of suggesting a seat limit increase.
  • Manual Only holds when your GitLab or Bitbucket account is linked to several Cursor accounts. Fixed an issue where Bugbot could still review your merge requests and pull requests automatically on GitLab and Bitbucket in that case. If Manual Only is on in any of the linked accounts, Bugbot now waits for you to trigger it.

Dashboard

  • Admin Clearer account list in the Bugbot dashboard. The coverage table now labels its first column Account, shows the GitHub account name for each installation, and marks personal accounts with a Personal badge. Accounts you manage are listed first, then organizations, then personal accounts. Bugbot's installation pages and MCP tools also open correctly from the Automations and Custom Agents pages.
  • Admin Sign in to team MCP servers from the Bugbot dashboard. When a team MCP server in Bugbot's MCP tools settings needs authentication, its status shows a Needs Auth — Sign in button that opens the server's sign-in page. Each click gets a fresh sign-in link, so a stale or used link no longer blocks the connection.
  • Open Settings from Bugbot pages with a keyboard shortcut. Press Cmd+, on macOS or Ctrl+, on Windows and Linux on Bugbot pages to open your Settings, the same shortcut the Cursor desktop app uses. It works on non-Latin keyboard layouts and does nothing while you're typing in a text field.

Fixes

  • Fixes to review reliability. Bugbot still posts its review when the Autofix pre-check fails, and skips Autofix for that run. On GitLab, a manual Bugbot command that hits a temporary user lookup failure now gets a reply explaining what happened instead of being dropped. Bugbot also handles its own outdated inline comments again, such as minimizing previous comments.

Agents Window

Chat and composer

  • Goal, queue, debug, canvas, and browser panels now open from pills above the prompt. Your conversation goal, queued follow-ups, Debug mode reproduction steps, canvas preview, and local dev server preview each have a pill above the prompt (Goal, Queued, Debug, Canvas, and Open Browser) that opens or minimizes its panel. Each panel opens by itself when there is something new, such as a new goal, your first queued message, or a ready repro step, and closing it keeps the pill.
  • Expand and copy Markdown tables in one click. Hover a Markdown table in an agent's reply or a Markdown file to show Expand table and Copy table buttons beside it. Expand table opens the table in a full-size view with wrapped columns and working links, and Copy table pastes as a real table into docs and chat apps, or as Markdown into plain text.
  • Download files from chat links. Right-click a file link in an agent's chat and choose Download File to save a local copy, with ~/Downloads as the default. It also works for files a cloud agent built in its workspace.
  • Mute toggle for video players. Inline and fullscreen videos in the Agents Window now have a Mute/Unmute button, and each new video starts muted or unmuted the way you left the last one.
  • Editing an earlier message keeps the agent's memory of the turns above it. Fixed an issue where editing an earlier message in a reopened or synced Agents Window conversation restarted the model from an empty history. The edit now resumes from that turn's state, and if that state can't be found, the edit fails and leaves your transcript intact.
  • Queued messages wait for Debug mode reproduction steps. When a Debug mode turn ends with reproduction steps, a follow-up you queued no longer sends right away and hides the steps. The Debug pill stays next to the queue until you click Proceed or Mark Fixed, or switch out of Debug mode.
  • Cloud agents can build canvases. Canvases a cloud agent creates are saved to the cloud and open as read-only previews in a Canvas tab, including from links in the agent's chat. For cloud agents, Canvas also shows up in the New Tab menu.
  • Pick the right element in Design Mode with the arrow keys. In Design Mode, you can now select a container even when a small child element sits on top: press ↑ to move the hover highlight to the parent element and ↓ to move back down, then click to select it.

Cloud agents

  • Fixes to cloud agent connections. Cloud agents reconnect on their own after your computer wakes from sleep instead of failing, and conversations and the agent list keep updating through network drops, proxy errors, and server deploys. A cloud agent whose machine went away reconnects its workspace when you send a new message.
  • Move a cloud agent into a local worktree. If a cloud agent works on a single repository you also have cloned locally, the environment picker in the status bar now lets you continue the agent in one of that repo's existing Cursor worktrees or in a new one, without touching your main checkout.

Navigation and layout

  • Closing and reopening tabs takes you back where you were. Cmd+W on a tab now returns you to the tab you were on before, even across split panes, and closing a split tile or pane puts focus back on the one you last used instead of the first or an empty one. Cmd+Shift+T reopens whatever you closed most recently, chat or app tab, including Agent, Environment, Changes, Desktop, Machine, and voice call transcript tabs.
  • Chat tabs that fit their titles, with agent details on hover. Chat tabs in the Agents Window now size to their titles and the tab strip scrolls when they don't fit. Hovering a chat tab, a standalone chat's title, or a tile header shows the same agent card as the sidebar, and moving between tabs moves the card without waiting again. Standalone chats and tiles now use the same tab strip as other panes, so you can double-click a tab to rename it, and with the sidebar hidden, holding Cmd (Ctrl on Windows and Linux) shows each agent's shortcut on its tab.
  • Drag tiles between splits. In a split, drag a tile's main chat tab onto another tile to move the tile there. A chat hidden from the sidebar also keeps its own name in its tile header.
  • Subagents and side chats open as tabs on their parent agent. Opening a subagent or side chat from a link, the command palette, or another agent now opens its parent agent with the subagent or side chat as a tab beside it, so you keep the parent conversation in view. A + button on the agent's tab strip opens its subagents as tabs. Back goes to where you were before the subagent opened, not to the parent agent. Subagents of automation runs still open on their own.
  • Reach every side chat from the side rail and new-tab menu. The side rail now lists your five most recently updated side chats, with the rest behind an "N more" menu, and when the rail is collapsed to icons a single Side chats button opens all of them. The new-tab menu's Side Chats and Plans submenus now list every entry instead of stopping at 10, and add a search box once there are more than 10.
  • Fixes to unread agents in the menu bar and Dock. Agents you mark as unread, or that finish while you're viewing another agent, now show in the menu bar and Dock menus without a restart. Unread agents take the first spots in the menu, with an "N more unread" entry when they don't all fit, and the badge counts only what the menu shows. A Project's row now lights up when one of its threads is unread.
  • Browser tabs reload when your dev server comes back up. A browser tab pointed at a local port that wasn't serving now reloads by itself when a terminal command or an agent's background shell starts a dev server on that port, so you don't have to refresh it by hand. You can also reload a tab with the new Reload Browser Tab command in the Command Palette or with Cmd/Ctrl+R.
  • File links in agent replies open faster. Hovering or focusing a file link in an agent's reply starts loading the file tab in the background, so the file is ready when you click.

Projects and workspaces

  • Clone a repo to run it on your machine. If a repository isn't cloned on your machine yet, its local row in the Run on menu now says "Not cloned". Pick that row to clone the repo. A notification shows progress while it clones.
  • Add local agents to cloud Projects. Drag a local agent onto a cloud Project in the sidebar, or use Add to Project, to make it a worker in that Project without losing its conversation. The agent keeps running on your computer through Remote Control, so Remote Control needs to be on for that workspace.
  • See what a Project's coordinator is working on. A Project's coordinator now reports a short status for its current step while it works, such as the subtask it is handing out or reviewing. The status shows live in the Project chat and in the fullscreen editor's agent overlay.
  • Easier-to-scan messages in Projects. In a Project's main conversation, your messages now show as flat, filled bubbles with no border, so they stand apart from the agent's replies. The fill follows your theme, and Tech Lead conversations use the same style.

Files, changes, and pull requests

  • Clearer pull request stacks and PR tab actions. In the PRs tray and the PR tab's stack navigator, closed PRs in a stack now fold into an "N closed" row, and a stack stays grouped even when only some of its PRs belong to the Project, with rows like "1 PR not in this project" marking the gaps. Merging an Origin PR that lands its open downstack now reads "Merge N PRs" and notes that it lands as one commit. The PR tab menu adds View on Web, follow-up pills such as Fix CI name the pull request they act on, and Fix with Agent on a Bugbot finding no longer clears your draft or resends the Bugbot report with your next message.
  • Open in Browser works for HTML files on remote and cloud machines. Right-click an HTML file in the file tree on a Remote SSH machine, cloud agent, or self-hosted worker and choose Open in Browser to preview a local copy. If it can't open, a message explains why, for example when the remote machine isn't connected or the file is too large.

Customization and settings

  • Adjustable chat width. A new Chat Width slider in Settings > Appearance > Agent Conversations sets how wide agent conversations can grow, from Narrow to Wide (600 to 1600px). Widen it to give tables and code more room, or narrow it for easier reading.
  • Fixed the translucent window tint on macOS when switching themes. On macOS, the Agents Window's translucent background tint now updates every time you switch between light and dark themes. Before, it could keep the previous theme's tint.

Plans and billing

  • Admin Clearer credit prompts and automatic top-up in the Buy Credits dialog. When a model needs credits to run, the usage notice now explains why and puts Buy Credits first if there's no plan to upgrade to. The Buy Credits dialog also has a Set Up row for automatic top-up, so you can save a top-up rule or turn it off without leaving the dialog. This applies to plans with a weekly usage allowance.
  • Admin Fixes to the Enterprise usage chart. On Enterprise plans, the Cumulative Spend chart in Settings > Plan & Usage now stays readable at large totals, with short axis labels like $1.5M, and a spend limit far above current spend shows in the legend instead of squashing the plot. Cached usage and chart data appear right away instead of a loading state, and an error or stale-data banner from a previous team or user no longer shows for the current one.

Fixes

  • Fixes to Computer Use on macOS. The Enable Computer Use permissions window no longer keeps reopening while permissions are missing. Computer Use also no longer gets stuck saying another Cursor window is installing it after a window quit or crashed mid-install.
  • Fixes to trays and pills above the follow-up input. Opening trays, pills, or prompts above the input no longer shifts the conversation, and interrupted tray animations no longer jump. Approval, error, and branch mismatch prompts can no longer be dismissed by accident with Escape, the environment setup popout keeps your edits when another tray covers it, and pills stay visible and show as selected only while their tray is open. Answering an agent's questions in the fullscreen editor no longer crashes the view.
  • Fixes to switching between agents. Switching back to a recent agent no longer reloads the chat, loses your scroll position, or flashes while it loads, and a hidden chat no longer captures your typing. A chat pop-out for a Project worker now stays open, and closes when its agent is archived or deleted.
  • Fixes to the sidebar and composer. Typing while an editor panel is fullscreen now goes to the prompt, the New Agent button no longer disappears from grouped sidebars, and long agent titles fade out instead of hiding the peer count. If starting a cloud agent from a sidebar draft fails, your prompt comes back instead of being lost. Stopping a new agent's first message keeps the browser panel and its tabs.
  • Fixes to file search, file tabs, and the Changes panel. File search and Quick Open stay responsive while you type, and pressing Enter opens the top result for what you actually typed. File tabs open faster on remote workspaces and refresh when a cloud agent edits the file. The Changes panel updates sooner after outside git commands, and a refresh no longer undoes a stage or unstage.
  • Fixes for errors when no editor is open. The Agents Window no longer throws errors when you open a file link, switch into the agent layout, or have unsaved files change while no editor is open. Changing the sidebar position and viewing a transcript for a conversation that is no longer loaded also no longer fail.
  • Fixes to cloud agent status and history. Cloud agents no longer get stuck on an outdated status, flicker back to running, or flip back to unread when a stale update arrives, and loading earlier messages reliably loads the previous page. Re-recorded videos and retaken screenshots show the new content, Project worker headers update right away, and only the coordinator's turn plays the finished chime. A cloud agent started on a picked self-hosted machine keeps that machine after refreshes and forks, and retrying a failed start from a saved environment keeps its repo configuration.
  • Fixes to agent transcript scrolling and loading. Scrolling long agent transcripts no longer makes messages jump or shows a blank band at the top. Opened agents appear sooner and show mid-turn messages right away. The transcript no longer shifts when a turn finishes.
  • Fixes to the @ mention menu. The @ menu no longer stays stuck loading when one source hangs, and the Branch row and past chats appear right away. For cloud agents, file mentions load as soon as you focus the input, typing stays responsive in large repos, and the slash command menu no longer goes empty while the cloud workspace starts. Editing a sent message now lists the same past chats as the follow-up input.
  • Fixes to agent chats. Charts in agent replies no longer clip tooltips or cut off long labels, Windows file path links open the file, and Copy Link on a PR chip no longer opens the PR. Cancelling a commit in Review Changes no longer goes on to open a pull request, and the new-agent model picker keeps the model and reasoning effort you just sent.
  • Fixes to Design Mode in the browser. The Design Mode pill no longer appears on a local browser tab when nothing is serving it or when its dev server runs in another workspace. Elements you pick stay on the right element after a hot reload.
  • Fixes to voice call cards. A voice call card for a call on another device no longer briefly shows an inflated duration when that call goes silent. It now shows as ended at the time it was last heard.
  • Fixes to plugin skills, chat sync, the compact composer, and the sidebar. Cloud plugin skills that failed to load now come back in the / and + menus when you reopen the menu or return to the window, and chats started or steered in another window no longer miss later turns. In a narrow compact composer, long model names no longer crowd out the text field, and the sidebar no longer lists the same folder twice when its path differs only in letter case.

Cursor IDE

Agent and chat

  • Auto-review stops risky deletes for your approval. Auto-review now asks before running a recursive, force, wildcard, or mirror delete whose target could resolve differently than written, such as an unquoted path with spaces, an unset variable, or a relative path after a directory change that can fail. Commands that match one of your block instructions now always show the approval card. With Run Everything on, agents now run without the sandbox unless your permissions.json restricts that mode.
  • Copied tables keep their formatting in docs and chat apps. Copy Table in the right-click menu of an agent response table now pastes as a real table in rich editors like Google Docs, Notion, and Slack. Plain-text destinations still get Markdown.
  • Smoother scrolling while an agent streams. When the agent chat follows a streaming reply, it now glides smoothly to new output instead of moving in uneven steps.
  • Agent Review fixes and better Fix in Cursor prompts. Fix in Cursor from a Bugbot report now drafts a prompt that lists the affected files and line ranges, even when code selections can't be attached. Agent Review tabs stay open when you open or close the chat Review tab or revert a plan checkpoint, and Agent Review shows again for users who had turned on the removed Beta "Hide Agent Review" setting.

Cloud agents

  • Cloud Agent links from the editor open in the Agents Window. Clicking an "Open in Cursor" link for a Cloud Agent while an editor window has focus now opens that agent in the Agents Window. Previously the editor tried to match the agent's repository and could show an "Incorrect Repository" dialog instead of opening the agent.
  • Cloud agents recognize out-of-memory kills. When a shell command in a cloud agent started from Cursor is killed because the machine ran out of memory, the agent is now told so, along with the memory limit, so it can explain the failure and retry using less memory instead of guessing.

MCP

  • More reliable MCP server sign-in. MCP servers whose OAuth apps reject redirect URLs that point to your own computer can now sign in through a Cursor-hosted callback page. Sign-in also completes when another Cursor window or app holds the local callback port, and no longer fails on large browser request headers.

Customization and settings

  • Easier hook correlation for shell and MCP calls. Shell and MCP execution hooks now receive a tool_use_id that matches the preToolUse event, so one script can follow a tool call from start to finish. Every hook now gets the .jsonl transcript in transcript_path when one exists, and CURSOR_TRANSCRIPT_PATH always matches it.
  • Claude Opus 5.5 and Sonnet 5.5 on AWS Bedrock. If you connect your own AWS Bedrock account, you can now pick Claude Opus 5.5 and Claude Sonnet 5.5 from the Bedrock model list in Cursor Settings.

Privacy and security

  • Admin Admins can restrict Agent to approved remote workspaces. A new managed policy, AgentRemoteWorkspaceAuthorities, turns off Agent in local Cursor windows and allows it only in remote workspaces. You can optionally limit it to remote authorities that match a comma-separated list of patterns, with * wildcards. If a user starts Agent in a local window or an unapproved remote workspace, they see a message that their administrator has disabled it.

Git

  • Clear error when Origin rejects a push for a file over 100 MB. When Source Control pushes to an Origin repository and the push is rejected because a file exceeds the 100 MB limit, the error now names the oversized file or files instead of telling you to pull first. It also explains how to fix it: track the file with Git LFS and remove it from every commit in the push.

Fixes

  • Admin Okta FastPass sign-in works in the built-in browser again. Signing in with Okta FastPass in the built-in browser no longer stops at "Okta FastPass requires network permission". The browser now grants the local network permission Okta Verify needs.
  • Fixes to agent tools. Agent edits to very large files no longer cut off the rest of the file, loose one-line replacements no longer splice the wrong text, and reads right after an edit in the same step return the updated file. The agent no longer hangs on unknown tool calls, and background shells survive idle chats. Also fixes a shell injection risk from crafted branch names when the agent gathers git diffs or looks up a branch's pull request.
  • Fixes to long-conversation summarization. Agent chats that hit the context limit now compact and continue instead of failing repeatedly at the limit, and summaries no longer drop facts from early in the conversation. The agent also keeps your latest request word for word after summarization, and near the limit it reuses a summary started in the previous turn instead of starting over every turn.
  • Fixes to MCP sign-in and settings. Connecting a Microsoft MCP server now shows the account picker, and edits to a project MCP server config show up in settings right away.
  • Fixes to remote connections and file watching. Connecting to an SSH, container, or tunnel remote no longer hangs when the connection drops during the handshake, and remote windows opened without a folder finish connecting. The explorer and editors keep picking up remote file changes after a reconnect, and absolute watcher exclude paths are now honored. The agent's branch display shows the new branch on the first refresh after a checkout.
  • Fixes to the built-in browser, settings, storage, and crash handling. New browser tabs no longer flash white, switching tabs no longer shows the previous page, and the Cursor Settings sidebar stays pinned with its bottom items reachable. Opening Cursor while a previous instance is still closing no longer discards local state, and sandboxed agent commands on Linux with AppArmor 4 can reach the network again. Repeated out-of-memory crashes no longer fill the disk with debugging data.
  • Fixes to Tab, Customize, and the Debug Console. Multi-line Tab suggestions now display correctly in files with Windows (CRLF) line endings. The "Use your skills with Cloud Agents" banner in Customize renders with its proper styling, and the Debug Console no longer throws an error when styles change before a debug session has been focused.
  • macOS no longer asks for Music or Photos access. On macOS, opening a folder that contains your Music, Movies, or Pictures folders no longer triggers Apple Music, Photos, or other access prompts. File watching and file search, including the @ menu, now skip those folders.
  • Admin Fixed Always Run on terminal approvals when your allowlist is managed. When a team policy or permissions.json sets your terminal allowlist, the agent's terminal approval card no longer offers Always Run. Before, choosing it had no effect, so the same command kept asking for approval.
  • Smaller local state files. Cursor now reclaims unused space in its local state databases when you quit, so they shrink over time instead of only growing. Cursor limits how long the cleanup runs, so quitting stays fast.

Web

  • More GitHub-style pull request search. Pull request search supports has:label and has:assignee. Searching review-requested:<user> or review-involves:<user> also matches review requests sent to that person's teams that you can see.
  • Merging no longer lands a pull request on another pull request's branch by accident. If a pull request targets the head branch of another open pull request it isn't stacked on, or of one closed without merging, merging is blocked with a message explaining where it would land. You can merge the parent first, retarget, reopen and stack it, or use Merge anyway on the pull request page.
  • Stacked pull request merge blockers say what to do. When a stacked pull request can't merge, the blocker tells you how to fix it, for example by naming the pull request to restack from, or the default branch to retarget onto when the base branch was deleted. A pull request with its own merge conflicts is no longer shown as only waiting on a conflict lower in the stack.
  • Undo a conflict resolution from the pull request timeline. When a conflict resolution is applied to a pull request, the Activity timeline shows it with the commit it created, and the most recent one has an Undo button that restores the previous head. The timeline also records who undid it and which commit was restored. For stacked pull requests, the button's tooltip shows how many pull requests in the stack the undo affects.
  • See and load every file on very large pull requests. When a pull request has more files than the Files view loads at once, the list shows how many files remain, counted by top-level folder and by added, deleted, modified, or renamed. On pull requests with more than 50,000 changed files, Load the next 50,000 files continues the list from where it stopped.
  • More accurate review filters in inbox sections. In inbox section filters, "Has 1 or more requested reviewers" counts only review requests nobody has answered yet, and "Has an active review" counts only open feedback from someone other than the author, ignoring dismissed reviews. A pull request requested from your team stops waiting on you once another team member reviews it, as on GitHub.
  • Custom code tours can cover up to 500 files. When you pick files for a custom code tour, you can select up to 500 of them instead of 100, and Generate no longer refuses a selection because the file paths are too long.
  • Link pull requests with # in comments. Typing # in a pull request comment suggests the repository's recent pull requests, so you can link to one without leaving the page. Before, the menu opened empty.
  • Filter a repository's commit history by author. The commits page has an author picker next to the branch selector. Choose a contributor or type an email address to list only that person's commits on the branch, and share the filtered view by its URL.
  • Simpler personal Slack notification settings. In Notifications settings, each personal Slack notification switch saves as soon as you toggle it and shows an error if the save fails, and you can connect Slack without leaving the page. Slack pull request cards have a Manage notifications menu, and review reminder digests link straight to your notification settings.
  • Admin The audit log records repository visibility changes. When someone makes an Origin repository public, internal, or private, your team's audit log records who changed it, the old and new visibility, and when, including when an agent's draft repository is published.
  • Clearer push errors and more reliable large pushes. When a push is rejected for a file over 100 MB, Git tells you how to fix it: track the file with Git LFS, and remove it from every commit in the push with git rm --cached or git filter-repo. Large pushes, including git push --quiet, no longer time out while Origin processes them, and pushes rejected by push rules show a message for each rejected ref instead of dropping the connection.
  • Fixes to the review inbox. The review inbox no longer jumps around while section counts load. Merged or closed pull requests stop showing up again in their old section, and agent runs on them no longer appear under Needs you. Inbox search picks up a merged pull request's corrected target branch within moments instead of up to an hour later.
  • Fixes to diffs. Files that start with a byte order mark now show their first line correctly in pull request diffs.
  • Fixes to conflict checks. While conflicts are still being checked, the merge box shows "Checking for conflicts…" and holds the merge instead of claiming the branch is clean.
  • Fixes to codebase pages and repositories. Load more in file and commit history keeps loading until new entries appear. Label names and descriptions are counted in characters, so emoji no longer hit the length limit early. Namespace names that collide with codebase routes, such as settings or apps, can no longer be claimed, and Public repositories no longer also show a Shared badge.

API

  • More ways to find pull requests and automate reviews through the Origin API. You can filter pull request lists to those carrying every label you name, and sort pull request search results by creation, updates, or relevance in either direction. Team API keys can submit pull request reviews in the repositories they're scoped to. See the Origin API changelog.
  • Admin Member emails and GitHub details in the Origin API require their own scopes. Origin API responses include actor emails only when the caller has the repository members read scope, and filtering pull requests by author email without it returns a permission error. Team reviewer rosters need the reviewers read scope, and GitHub IDs and mirror details need the mirror read scope. If your app relies on these fields, request the scopes and have an admin approve them; see the Origin API changelog.
  • Origin MCP tools accept more argument styles and link to comments. Origin MCP tools accept GitHub-style argument names such as owner, repo, and pull_number, a repository given as owner/name, and enum values in any case, instead of rejecting the call. The pull_request_read tool's comments and threads views include a url that opens that comment or thread on the pull request page.
  • Fixes to Origin API responses. API reads and webhooks no longer fail when an actor's account was deleted, and pull requests, reviews, comments, and reactions show their real author, including service accounts. Paging through change lists, change search results, and comparison files no longer skips or repeats items. See the Origin API changelog.

Projects

  • Agents can message and read agents you @-mention. When you @-mention another agent to a Project coordinator, the coordinator can read that agent's transcript and send it messages. It can also message its own parent coordinator, and the agent that receives a message can reply directly. Access follows yours: the coordinator reaches only agents you mention and can still open.

Agents

  • Subagent progress shows up live in the parent agent. When a Cloud Agent hands work to a subagent, the subagent's tool calls and messages now stream in under the parent's task as they happen, instead of appearing only after the subagent finishes.
  • Agent links in summaries now open the right page. When a cloud agent's summary links to another agent, the link in Slack, GitHub, GitLab, and Bitbucket now opens that agent's page on cursor.com instead of breaking. On the web project board, agent links in coordinator files now open the linked agent, including its Changes tab when the link points there.

Environments

  • Comments and trailing commas in environment.json work everywhere. Cloud Agents now fully support // and block comments and trailing commas in .cursor/environment.json, wherever the file is read. Repos with a commented environment file no longer get the "set up your environment" prompt, and their snapshot and network settings now apply instead of being silently ignored.

MCP

  • Cloud Agents handle MCP rate limits without retry loops. When an MCP server rate-limits a tool, the agent pauses that tool for as long as the server asks, keeps using the server's other tools, and is told when it can retry. Repeated Slack rate limits now cap the wait at about a minute.

Self-Hosted Machines

  • Project workers on a named Self-Hosted Machine use that machine's repository. When a Project coordinator starts a worker on one specific Self-Hosted Machine, the worker now works in the repository or folder registered on that machine instead of the Project's repository. Coordinators without a repository of their own can now place workers on a named machine too. Targeting a pool instead of one machine returns an error asking you to name a machine.

Slack

  • Admin Choose which team your Slack agents run under. If you belong to more than one paid team, you can now pick the Cursor team that agents you start from Slack run and bill under. Choose it when you link Slack, or change it later with the Team for Slack agents setting in Cloud Agents settings. Follow-ups in a Slack thread keep running under the agent's original team, and a team pinned by your Slack workspace takes precedence.
  • Slack previews for agent artifact links. Links to a cloud agent's artifacts, including public artifact shares, now unfurl in Slack with the file name and type, plus the length for videos. Previews only appear when the person who shared the link has access to the artifact, and are skipped in channels shared outside your workspace.

Fixes

  • Fixes to agent runs and follow-ups. A message sent after pressing Stop now runs instead of staying queued, and answered questions stay answered after a reload. A run no longer shows a PR from another branch as its PR link.
  • Fixes to repository access and tokens. Cloud Agents refresh GitHub tokens during long runs instead of hitting expired tokens mid-run. Repository access errors at start now name the specific repository that is empty, denied, or needs connecting.
  • Fixes to the Cloud Agents web app. Keys like Ctrl+K and Ctrl+F in the agent terminal now go to the shell instead of triggering page shortcuts. The Connect GitHub card names the repository an agent needs access to.
  • Fixes to MCP tool calls. MCP tool calls recover after a server restart instead of failing, and invalid-argument errors now show the specific reason. Google Calendar time-range searches return only events in that range, and Gmail searches of Spam and of very recent mail no longer come back empty.
  • Fixes to Self-Hosted Machines. Agents on any-repo pools no longer pick up untracked rules or instructions left in a checkout by an earlier run. When no worker matches your labels, the error now lists the requested labels.
  • Fixes to Slack launches. Changing or resetting your default model and settings from Slack now works for accounts linked under a migrated Slack workspace. When an agent started from Slack fails, the ⏳ reaction is now replaced by ❌ instead of staying next to it.
  • Admin Fixes to Remote Control. When Remote Control is turned off, your phone now shows it as disabled instead of a generic unavailable error. Members of enterprise teams that aren't linked to an organization can no longer turn on Remote Control until an organization admin allows it.

Integrations

GitHub

  • Admin Pull request details stay limited to people with GitHub access. Cursor shows pull request details only to people whose GitHub account can read the repository, even when a teammate already loaded them. Everyone else gets an access error.
  • Clearer errors when GitHub blocks a merge method. Merging a GitHub pull request from Cursor with a merge method the repository doesn't allow now fails right away with GitHub's own reason, such as "Squash merges are not allowed on this repository." Before, Cursor retried and then showed a generic can't-merge-yet message.
  • Fixes to GitHub status and setup. Pending commit-status checks now show as running instead of queued, Actions log excerpts no longer hang on redirects, and @cursor commands from people whose GitHub username contains "cursor" are no longer ignored. GitHub Enterprise Server app installs now land on the team that started setup.

GitLab

  • GitLab repository lists and merge request lookups load reliably. The Bugbot repository list for large GitLab groups loads with the projects fetched so far instead of failing on a timeout, and finding the merge request for a branch now works for GitLab instances reached through AWS PrivateLink or an egress proxy.

Linear

  • Full Linear team names on hover. Hover a truncated team name in the teams list on the Linear integration page to see the full name.

Source control

  • Steadier self-hosted source control connections. Cursor retries a read request once when a connection to GitLab self-managed or Azure DevOps drops briefly. Over AWS PrivateLink, GitHub, GitLab, and Bitbucket requests also retry once when the server closes the connection.

Dashboard and admin

Plugins and MCP

  • Admin Control which MCP and plugin tools your team can use. Admins can turn Team list enforcement on or off from the Plugins page and limit a URL MCP server's or plugin's tools from its detail dialog, picking allowed tools from a live preview. The MCPs tab has an Allowlist rules editor that marks rules that apply only while Team list is on, and each server shows its access status, such as Read-only or Cloud only, with a tooltip that explains why the IDE and CLI block it. See MCP server trust management.
  • Admin Browse and add plugins and MCP servers from one list. The Plugins & MCPs page shows installed plugins and MCP servers together in one searchable table, with an Add button for the selected scope. Clicking a plugin opens its details in a dialog instead of a separate page.

Members and groups

  • Admin Set soft or hard spend limits for a group's members. In organization group settings, admins can set a member's monthly spending limit to No limit, Soft limit, or Hard limit. A soft limit doesn't block members and can be paired with pace alerts. A hard limit blocks members once their monthly spend reaches the amount. See group spend limits.
  • Admin Groups show their slugs, label synced groups correctly, and all appear on large teams. Team and organization group lists and group pages show each group's slug under its name. On the team Groups tab, groups synced from your identity provider are labeled Synced instead of Manual, and Move member no longer offers synced groups as destinations. Teams with more than 200 groups now see all of them on the Groups tab, in the Move member list, and in the invite group picker.
  • Admin Pending invites close when someone joins your team. When someone joins your team in any way, such as SSO, domain join, SCIM, or a direct add, their other pending email invites to it close automatically, and an old invite link still takes them to the team. When someone joins through a copied invite link, the audit log records the member who copied the link as the inviter.

SSO and SCIM

  • Admin Turn on SSO auto-enrollment from either SSO page. The SSO auto-enrollment switch now also appears on the older team SSO settings page, and both SSO pages show a newly finished SSO setup as soon as you return to the tab. When your identity provider marks a SCIM user inactive, Cursor removes their team seat but keeps their SSO membership, so reactivating them doesn't require adding them again. See SCIM user management.

Agent settings

  • Admin Control where Project agents can run. The new Automatically allow Projects to place their work switch in Cloud Agents security settings decides whether a Project's agents can start work on any machine, environment, or repository the Project owner can reach without asking first. When it's off, agents ask for approval before they use a new one. It starts off for Enterprise teams and on for other teams. See Cloud Agents security settings.
  • Choose which team your Slack agents run under. If you belong to more than one paid team, the Team for Slack agents setting in Cloud Agents settings picks the team that agents you start from Slack run, bill, and use settings under. The setting is locked when your Slack workspace is restricted to one team. See Team for Slack Agents.

Automations

  • Mine, Team, and Explore tabs on the Automations page. Mine lists the automations you own, Team shows your team's automations alongside agents from Cursor, and Explore holds the template gallery. Team admins open on Team by default, and everyone else opens on Mine.
  • Admin Security Reviewer and Vulnerability Scanner have their own pages. The From Cursor automations list shows separate Security Reviewer and Vulnerability Scanner cards in place of the single Security card. Each page lists only its own agents and recent runs, its New button creates that type directly, and flagged vulnerabilities appear on the Vulnerability Scanner page. Old Security links redirect to Security Reviewer. See Security Review and Vulnerability Remediation.
  • Start Rollouts checks yourself when a deploy isn't recorded. If a merged change goes live without a recorded deployment, select Run checks on the change's page in Rollouts, pick the environment and services, and Rollouts starts checking it. Deferred checks show as Deferred with the reason, and closing an issue now asks for a reason, such as Fixed or Duplicate. See Rollouts.

Usage and integrations

  • Admin Filter usage by SDK. The Usage pages for teams and organizations have an SDK option in the client filter, so you can see spend and requests from agents run through the Cursor SDK on their own. Now available to everyone.
  • Admin Connect SharePoint for your whole team. Admins can connect SharePoint from the Integrations page: choose delegated or application permissions, enter your Microsoft tenant, grant admin consent, and optionally allow write access. If setup stops partway, members who add the SharePoint plugin are told to ask an admin, and admins get a link back to Integrations to finish.

Security and privacy

  • Admin Bitbucket connection changes in the audit log. The audit log records when a Bitbucket Cloud or Bitbucket Server connection is connected, disconnected, registered, or deleted, and when its token is set or cleared, with the provider and host. Filter for them with the new Source Control Connection event type. See audit logs.
  • Admin Clearer privacy mode names. Privacy settings now read "Privacy Mode (no training)" and "Legacy (no storage)". When you invite teammates from a personal plan, a new "Share Cursor Analytics" checkbox sets the new team's privacy setting, and Privacy Mode stays on unless you check it.
  • Admin Control Grok dictation and private chats. Team admins can turn dictation on or off for their members from the Grok tab. Organization admins also get a Features section there with switches that turn off dictation or private (temporary) chats for every team in the organization, so all chats are kept for compliance exports.

Fixes

  • Fixes to dashboard layout and display. Team settings opens on the first click, and plan reset and billing cycle dates show in your local time zone instead of UTC. The Usage page shows how far you are over your monthly limit instead of $0 remaining, and the Overview setup checklist shows the steps that loaded instead of failing the whole card. Long names and emails truncate with the full text on hover, and pages no longer flash stale content or jump while loading.
  • Fixes to the dashboard on phones. The navigation drawer can be swiped left to close, the Members table and Usage filters fit the screen, and the Usage date picker no longer widens the page. The team name editor wraps and stops at 200 characters, and the invite email field no longer autocapitalizes or autocorrects.
  • Fixes to the Automations page layout. The Automations list no longer leaves a gap between each automation's title and author, the "Showing N" count is the right size with proper spacing, the filter tabs line up with the page title, and shared GitHub and GitLab triggers use a generic git icon.
  • Fixes to Rollouts settings. Clicking Authenticate on an MCP tool no longer wipes an existing shared connection, and already-connected servers refresh instead of opening a sign-in window. The data sources list shows every row instead of a small scrolling box, and notification help text wraps on mobile.
  • Fixes to environments and secrets. An environment's page shows history for that environment only. Double-clicking Duplicate creates one copy instead of two, and choosing All repositories in the Edit Secret dialog saves the secret for every repository.
  • Admin Fixes to integrations and MCP authentication. Registering a GitHub Enterprise Server app attaches it to the team you selected, and the existing-apps list shows that team's apps. Authenticating an MCP server for a service account returns you to the settings page you started from, and the private inference assignee picker loads groups from every linked team. Paging through the Origin repositories card no longer hides the card or shows the wrong page when a page change fails.
  • Admin Fixes to account sessions and deletion. When an account's email changes to a different address through your identity provider, SSO, or directory sync, existing Cursor sessions on that account are signed out. Deleting an account that was previously merged with another account no longer fails.

Automations

Automations page

  • Mine, Team, and Explore tabs on the Automations page. The Automations page now opens on Mine, which lists the automations you own across your personal account and teams, alongside popular templates. Team shows your team's automations, and Explore holds the full template gallery. Now available to everyone. Team admins land on Team, and users without a team see Mine and Explore.
  • See each automation's tools at a glance. The Tools column on the Automations page shows each tool as a pill with its icon and name. When a row has more tools than fit, a +N pill shows how many more there are.

Triggers

  • Microsoft Teams automations now fire on every channel post. Automations that listen to a Microsoft Teams channel now run on regular posts in that channel, not only on posts that @mention Cursor.
  • Failed Slack-triggered runs are marked on the trigger message. When an automation started from a Slack message fails, that message now gets a red X reaction in place of the hourglass. Before, the hourglass disappeared and nothing showed that the run had failed. The channel still gets no failure post.

Security agents

  • Vulnerability Scanner follows your custom instructions and files follow-ups. Team Vulnerability Scanner automations now use your custom instructions to shape what they scan. After a scan, they carry out the follow-up actions those instructions ask for, like filing a ticket for each new finding, through the automation's MCP servers, and link what they filed in the finding's Slack report.

Automation runs

  • Clearer reasons when a run fails. Run history now tells you whether to retry or contact your admin when your organization's policy check blocks a model request or is unavailable. Runs that fail because no self-hosted worker matches their labels now name those labels.

Fixes

  • Fixes to the Vulnerability Scanner. Findings recorded by the Vulnerability Scanner show up again instead of appearing empty. Repositories listed as host/owner/name without https:// now scan instead of failing. Teams that turn Security off in their managed automation settings no longer get scheduled scanner runs.
  • Fixes to Change Monitor setup. Turning off Change Monitor now stops its agents, and asks you to try again if they don't stop. Saving repositories the installation can't see, or that are missing a webhook, now shows the specific setup error instead of a generic one. Nested GitLab paths now ask for the full gitlab.com/<group>/<project> form, and fully qualified GitLab projects are watched correctly.
  • Admin Fixes to the automation editor and integrations. Your GitHub or GitLab connection no longer shows as disconnected when you belong to several teams or use self-hosted GitHub Enterprise or GitLab. The editor's MCP server options now list the selected team's servers. Disconnecting a team's PagerDuty integration now removes the team's connection, and "Configure here" links in Security Reviewer comments and Slack messages open that agent's page under From Cursor.
  • Fixes to the Automations page layout on web and mobile. On narrow screens, the Automations list now scrolls sideways instead of cutting off columns, and the row actions stay pinned on the right. Spacing and alignment across the page header, tabs, toolbar, and list rows are fixed.

APIs

  • Admin Clearer error for Azure DevOps repos in the Bugbot repo update endpoint. POST /bugbot/repo/update now returns a 400 for dev.azure.com and visualstudio.com repo URLs, explaining that Azure DevOps repo settings are managed from the Cursor dashboard, instead of telling you to install the Cursor GitHub or GitLab app.

Security and approval agents

Security agents

  • Admin Separate Security Reviewer and Vulnerability Scanner pages. Under From Cursor in the dashboard, the Security card is now two cards, Security Reviewer and Vulnerability Scanner. Each opens a page that lists only that type of security agent and its recent runs, with a New button that creates that type and an icon that shows whether it's enabled. Old Security links redirect to Security Reviewer.
  • Incomplete reviews no longer auto-approve. When part of a security review fails, Security Reviewer no longer treats the result as clean, so it doesn't auto-approve the PR. If no reviewer finishes, it posts no findings and leaves existing threads alone. Earlier findings stay open when their follow-up check times out or fails.
  • Accepted findings aren't posted again on every push. If you resolve a finding without fixing it, for example to accept the risk, a later push that finds the same issue no longer opens a new comment. A different issue on the same lines still gets its own comment, and a fixed issue that comes back is reported again.
  • Failed large reviews finish sooner. When every reviewer on a very large security review fails, the review now stops early instead of retrying until its budget runs out.

Approval agents

  • Clearer approval requirements on pull requests. PR Routing & Approval checks, comments, and the reviewer hovercard now say exactly what clears review, such as "Needs 1 approval from any human reviewer" with suggested reviewers, one approval from any of a set of owners, or approvals from all named owners. Auto-approved low-risk pull requests no longer show a leftover approval requirement.

Models and Cursor Router

  • New model: Claude Haiku 5.5. Claude Haiku 5.5 is now available in Cursor as an opt-in model. Turn it on in Cursor Settings > Models.
  • Fixes for Mistral models. Mistral requests that failed with a validation error now complete, and the model's reasoning shows up as thinking, separate from its answer.
  • Vendor names for Alibaba and Mistral models. The model list now labels these models with their vendor.

Bitbucket

  • Admin Clearer Bitbucket Data Center support. The Bugbot dashboard now names Bitbucket Data Center organizations by their server hostname and labels them Bitbucket Data Center instead of bitbucket.org. Bugbot no longer offers Autofix on Bitbucket Data Center, and its review footnote no longer suggests switching to direct push there. Bitbucket Data Center requires a Teams or Enterprise plan.

Reviews

  • Local reviews say what they skipped. When you run Bugbot from your agent or run a security review on local changes and the diff is too large to review in full, the result now says how many changed files were covered and lists the ones that were not reviewed, so you can review them next. Findings in those skipped files are no longer discarded as unrelated. If every reported finding is filtered out, Bugbot says the review was not clean instead of reporting no bugs.
  • Admin Manual Only stops automatic learning from merged pull requests. With Bugbot set to Manual Only for a repository or in your personal settings, it no longer learns rules automatically when a pull request is merged or closed, and that learning no longer counts toward your Bugbot usage. @cursor remember comments still teach it rules.

Dashboard

  • Admin Consistent daily counts in Bugbot analytics. The Bugbot analytics dashboard now counts pull requests reviewed per UTC day, so a day's numbers no longer change when you move the date range. The PRs reviewed total counts each pull request once for the whole range, and the date picker is labeled UTC. The review analytics API now counts only reviews inside the requested range, so it matches the dashboard when you request all PR states.

Fixes

  • Admin Fixes to reviews and Autofix. With run-once enabled, a commit pushed while a review is still running now gets reviewed instead of skipped, and incremental reviews keep your BUGBOT.md rules even when the new push has no diff. Review comments no longer say a cloud agent was started to fix issues when it wasn't; they say the agent failed to start or that no issue met your Autofix severity threshold. Reviews on repositories granted to an Enterprise team now run under that team instead of the installation owner.
  • Fixes to PR summaries. If a PR summary hangs or fails, Bugbot no longer leaves a stale "generating a summary" placeholder in your PR description or comment. It removes it or shows "Unable to generate summary." Images, videos, and files you paste under the summary are no longer wiped when Bugbot refreshes it.
  • Fixes to the Bugbot dashboard and integrations. Connecting a self-hosted GitLab instance from the Bugbot dashboard now links it to the right team, and the repository picker during app install now shrinks to fit the window, so its search field no longer opens off-screen. Turning on Bugbot for an Azure DevOps repository after the enterprise app was re-consented now recovers automatically instead of failing. Per-repository Bugbot settings are cleaned up when a GitHub App is uninstalled or repositories are removed.

Cursor mobile

Review

  • Source and base branches in Review. The pull request Review screen in Cursor for iOS shows the source branch and the branch it merges into under the title, and the base branch follows the pull request's current target. VoiceOver reads full branch names even when long names are cut off on screen.

Fixes

  • Chat and review display. Text references from cloud agents show as cards instead of raw markup when their tags span multiple lines. When a cloud agent's diff is too large to show, the Review screen says it has too many files or some files are too large. Visualizations that can't render on mobile tell you to open them in the Cursor desktop app.

Agents Window

Chat and composer

  • Cmd+F reliably searches the chat. Cmd+F now opens find in the chat transcript whenever the chat or prompt has focus, or focus is anywhere outside the editor panel and Customize. File, diff, terminal, and browser find open only when you're working in that panel, and the fullscreen conversation tray now has its own find bar. Chat find also matches only text you can see, no longer flashes "0 of 0" while it searches, and uses the same yellow highlights as diff and browser find.
  • Move through long conversations from the keyboard. Press Cmd/Ctrl+Up or Cmd/Ctrl+Down to jump between your messages, and Cmd/Ctrl+Shift+Up or Cmd/Ctrl+Shift+Down to jump to the top or bottom of the conversation. The shortcuts also work while the prompt has focus, as long as it's empty.
  • Reference pull requests with #. Type # in the prompt box to pick from your pull requests, including the ones the current agent opened, and add one to your message as context. This also works in follow-up question inputs and when you edit a sent message.
  • Minimize or remove the plan above the prompt. The plan card above the prompt now has a Minimize button that collapses it to a Plan pill, so it stays out of the way while you keep chatting. Right-click the Plan or Canvas pill and choose Remove to clear it. The plan card also takes priority over follow-up suggestions and opens once another card closes.
  • Approval prompts explain why they appear. When an agent asks to edit or read a file outside your workspace, or to edit a protected config file, the approval card now says why, for example "This edits a file outside your workspace." Read approval cards show the full file path with an Allow button.
  • Queued messages no longer get lost. Messages you queue while an agent is working are no longer dropped when a turn ends.

Cloud agents

  • Drop local files into prompts for SSH, WSL, and dev container agents. When you drop, paste, or attach a file or folder from your computer into the prompt of an agent running on a remote machine, the Agents Window copies it to that machine and mentions the copy, so the agent can read it. Copies go to a separate uploads folder outside your workspace, so they never show up as workspace changes, and slow copies show a progress notification.
  • Drop a folder onto a cloud agent prompt to attach its files. Dragging a folder onto the prompt of a cloud agent now attaches the files inside it, skipping hidden files. Up to 20 files are attached per drop, and a notice tells you when a folder had more and only the first 20 were added.
  • Slack mentions are readable in cloud agent prompts. When a cloud agent starts from Slack, people mentioned in the prompt now show as @Name instead of raw Slack user IDs. Clicking a mention opens that person in the Slack app if it's installed, or in Slack on the web otherwise, and Find matches the names as shown.
  • New cloud agents use your repo's default environment. When you pick Cloud for a new agent, by clicking it or cycling with Cmd+;, the agent now runs on the repo and Cursor picks its environment. Before, it could preselect another teammate's saved environment. If an agent starts with a warning, such as falling back to the default image without your environment's setup, the environment status row in the chat shows the warning, counts it, and opens on its own.
  • Continue on Cloud shows saved environment names. Continue on Cloud options now show each saved environment's name, so you can tell apart two environments on the same repo.
  • Cloud Agent links open the right agent. "Open in Cursor" links for Cloud Agents from cursor.com/agents, Slack, Linear, Teams, or Copy Deep Link now open that agent in the Agents Window, even if you haven't opened it before, and even when the IDE has focus. Previously the link could drop you on a new chat. If the agent can't be found or you don't have access, you now see an error.
  • Side chats now work on cloud agents. You can open a side chat from a cloud agent with /side, the + menu, or by selecting text and choosing Add to Side Chat, the same way as on local agents. The side chat runs as its own cloud agent, so you can ask a question without interrupting the main run. Side chats aren't offered inside Projects, which use threads instead.

Projects and workspaces

  • Edit Workspace keeps your chats when you rename a project. Renaming a multi-root project in Edit Workspace now moves its local chats and drafts to the renamed project instead of leaving them behind. You can also remove folders until only one remains.

Files, changes, and pull requests

  • Copy an agent's pull request link from the Copy menu. When an agent has opened a pull request, the agent's Copy menu in the panel and sidebar now includes Copy PR Link. It copies the link for your preferred review destination, such as GitHub or Graphite, and confirms with a toast.
  • Clearer pull request status and faster PR actions. Pull requests whose checks are only waiting on someone's approval now show under a yellow Action Required group instead of CI Failing, and mergeable PRs read Ready. In the PRs tray, merged PRs in a stack fold into an "N merged" row you can click to expand, and middle-clicking a PR row in the tray or apps rail opens it in your browser. In a PR tab, press Cmd+Enter (Ctrl+Enter on Windows and Linux) to submit a reply in a discussion thread.

Navigation and layout

  • Switch command palette filters with the arrow keys. When the search box in the command palette is empty, press Left or Right to move between the filter pills, such as All, Agents, Files, and Actions. The footer shows the arrow-key hint when this works.
  • Drag subagents and chat links into a split. You can drag a subagent from the subagent tray, or a chat link in a conversation, onto a conversation tile to open it in a split pane. Top, bottom, left, and right drop zones now work more reliably on wide or narrow panes.
  • Clearer syntax colors in the Cursor Light themes. Cursor Light and Cursor Light Colorblind now give each kind of code its own consistent color, so code is easier to scan. Types are underlined in blue, properties are blue, decorators and macros are orange, and parameters are italic.
  • Press Escape to show or hide the Apps panel. When focus isn't in a text field and no dialog or menu is open, Escape toggles the Apps panel, and exits it first if it's fullscreen. Escape in the prompt also dismisses informational trays above the composer, and closing a dialog, menu, or fullscreen video with Escape no longer toggles the Apps panel too.

Customization and settings

  • Buy credits and manage your plan from Plan & Usage. On plans with a weekly usage allowance, hitting the weekly limit now shows a notice with Buy Credits and Upgrade Plan. Buy Credits opens an in-app purchase dialog, and the block lifts once the purchase goes through. Plan & Usage settings also show a Credits section with your prepaid balance, pending top-ups, and automatic top-up, plus a line showing when your weekly usage resets.
  • MCP tools that require confirmation always ask in Auto-review. Fixed an issue where the Auto-review classifier could approve MCP tool calls that the server marks as requiring confirmation, such as merges, without showing you an approval prompt. These calls now show the approval card unless you use Run Everything or have allowlisted the tool.
  • Fixed Windows hooks that point to extensionless scripts. On Windows, a hook command that names a script without an extension, such as a plugin's scripts/run, now runs the matching .ps1, .exe, .bat, or .cmd file next to it. Before this fix, these hooks opened a new console window, never received their input, and reported success without doing anything.
  • Fixed repository access rules for people on more than one team. Admin repository allowlists and blocklists now come only from your active team, so another team's allowlist no longer blocks repositories you should be able to open. Switching teams applies the new team's rules right away.
  • Turn Cursor Tab on or off from any file tab. The "..." menu on a code file's tab now always includes a Cursor Tab switch, so you can turn Tab completions on or off without leaving the editor. Your keybinding style, such as Vim or Emacs, also stays selected in that menu instead of resetting to Default after an update.
  • Settings now explain a missing menu bar icon on macOS. If macOS has Cursor turned off under System Settings > Menu Bar > Allow in the Menu Bar, the Menu Bar Icon setting now says so and offers an Open Settings button to turn it back on. Cursor also restores the icon automatically when an older hidden-icon preference was keeping it out of view.
  • Clearer explanation of a fixed on-demand limit. With a personal fixed on-demand limit, Settings > Plan & Usage now explains that requests stop at the limit and that overage from requests already running isn't billed unless you raise the limit that billing cycle. It previously said usage past your limit would be billed later. Team-managed limits keep their existing wording.

Fixes

  • Fixes to the Agents tray and subagents. Back and Forward now step into and out of subagent previews opened from the Agents tray, including previews opened from a side chat. The collapsed Agents pill no longer shows a stopped subagent as needing attention, and local subagents that failed no longer stay stuck in the tray (you can still open them from their Task block). Cloud subagent Task cards now show the subagent's name and progress even when the subagent starts after the card appears.
  • Fixes to cloud agent creation and follow-ups. A prompt that fails to start a cloud agent comes back so you can retry, and back-to-back follow-ups arrive in the order you sent them.
  • The last messages in a chat no longer hide behind the pending question panel.
  • Unsent drafts longer than 100,000 characters now survive a reload instead of coming back empty.
  • An image dropped into the follow-up composer now attaches once, not twice.
  • Typing a follow-up no longer crashes the agent panel.
  • Answered agent questions no longer reopen.
  • With text selected in a conversation, Cmd+L adds it to that conversation, including side chats.
  • Fixes to trays and pills above the prompt. Pills above the prompt no longer jump or flicker when a tray or banner opens or closes, and swapping one tray for another now crossfades in place without shifting the layout. The plugin install confirmation now sits above other suggestions with View Details beside the confirm button, and the plugin-added notice can be dismissed with Escape.
  • Fixes to agent turn reliability. Provider decode errors are now retried automatically.
  • Fixes to MCP server sign-in. Connecting to Microsoft Entra-backed MCP servers such as Azure DevOps no longer asks for consent on every sign-in, so tenants where an admin has already granted consent stop landing on "Approval required."
  • Fixes for Windows, git, and non-QWERTY keyboards. Plugins now load with git 2.26 and 2.27, and on Windows the minimize, maximize, and close buttons no longer disappear after you close a modal. On layouts like AZERTY and Dvorak, shortcuts on the physical Z and Y keys, such as Cmd+W, no longer trigger undo or redo.
  • Fixes to branches and git actions. The automation branch picker shows the default branch right away, and pull requests the local agent opens from its shell are now detected.
  • Fixes to cloud agent git actions. Fix Merge Conflicts stays in its loading state while the cloud agent works, and Checkout Branch and Fast Forward Branch no longer appear for cloud agents whose checkout can't be changed.
  • Fixes to the sidebar and project picker. Choosing Rename from an agent's right-click menu now reliably opens the rename field, and the sidebar footer shows your account name and avatar right away at startup. When every agent is pinned or in a project, a Chats header keeps the sidebar customize menu reachable, and the No Repo section no longer offers a Remove from Sidebar action that did nothing. In the project picker, long folder paths are shortened in the middle so the folder name stays visible, and row controls appear when you highlight a row with the keyboard.
  • Unread badge matches the sidebar. While the Agents Window is open, the Dock badge and menu bar unread count no longer count cloud agents that aren't in the sidebar.
  • Accessibility fixes. Screen readers now announce the highlighted item and its position in composer menus like slash commands, name agent tiles "Panel 1", "Panel 2" with a "Panel actions" menu, and treat collapsible sidebar section headers as expandable buttons. Buttons in the chat transcript show a focus ring when you reach them by keyboard, and dialogs focus the primary button more reliably when they open, so Enter confirms.
  • Fixes to sign-in, screenshots, and terminal links. Enterprise members behind flaky corporate proxies are no longer signed out after a single failed team membership check. Agent screenshots of a hidden or minimized window time out after 5 seconds instead of hanging, and finished terminal commands no longer auto-open a browser for a server started in a different terminal.
  • Fixes to voice chat. An earlier spoken request that got no reply no longer stays open for the rest of the call once a later request finishes. When the agent is waiting on you, the voice explains once that it can't take a new request instead of repeating itself. Hiding the chat no longer hides the floating call controls, so Hang up stays on screen.
  • Fixes to agent questions and tool calls. Questions and plan approvals the agent asks after a dropped connection are no longer answered with an earlier reply. Built-in tool calls with slightly malformed arguments now run instead of failing, and a native code chunker that fails to load now only disables semantic search. The transcript also holds its position when you scroll past the bottom while new output streams in.
  • Fixes to stability and local storage. Switching between Private Inference and Cursor-managed models now stays switched after the restart that applies it. Opening the Agents Window while a previous instance is still closing no longer discards local state, and repeated out-of-memory crashes no longer fill the debugging-data folder with gigabytes of files. Okta FastPass sign-in works in the built-in browser again.
  • Fixes to the Agents Window. Pull request status for agents keeps showing the last known state when GitHub rate limits requests, instead of failing to load. The pinned Build card no longer shows content through it under translucent custom themes. Closing Settings returns straight to the same conversation with its transcript and composer intact.

Cursor IDE

Plans and billing

  • Admin Plan & Usage shows your plan and prepaid credits. On Premium, Plus, Super, and Ultra plans, Cursor Settings > Plan & Usage shows your plan by name and your usage reset date. Individual accounts on these plans also get a Credits section with the prepaid balance, pending top-ups, and the auto top-up rule, plus buttons to add credits or change the rule.

Agent and chat

  • Approval prompts now tell you why the agent is asking. When the agent asks to read or edit a file outside your workspace, or edit a protected config file, the approval card shows the reason.
  • Clearer errors when image generation fails. When the agent can't generate an image, it now tells you why instead of showing a raw request error. A content-policy block suggests describing the subject in your own words, and a rate limit or provider outage asks you to try again.

Cloud agents

  • Slack mentions in cloud agent prompts show as readable names. When you open a cloud agent started from Slack, people mentioned in its prompts appear as @Name instead of raw Slack user IDs, and find in the conversation matches the names you see. Clicking a mention opens that person in Slack.

Editor

  • AI features recover on their own when the extension host hangs. If the extension host that powers Cursor's AI features stops responding for about a minute, Cursor restarts it automatically, so you no longer need to reload the window. It also keeps restarting a crashing host with increasing delays, and skips the automatic restart while you debug extensions.

Privacy and security

  • Privacy mode now blocks automatic crash debugging uploads. With privacy mode on, Cursor no longer uploads debugging data or emergency memory profiles on its own after a crash or out-of-memory event, even if you turned privacy mode on mid-session.

Fixes

  • Fixed Windows hooks that point to extensionless scripts. On Windows, a hook command that points to a script without a file extension now runs the matching .ps1, .exe, .bat, or .cmd file next to it. Before, these hooks opened a stray terminal window and reported success without running.
  • Fixed transparent images turning black when attached to chat. When Cursor shrinks a large image you attach, transparent areas now stay transparent.
  • Fixed Investigate Error in Chat taking over its shortcut. Cmd+Shift+D now runs Investigate Error in Chat only when the cursor is on an error, warning, or AI lint. Elsewhere, the shortcut goes to its other binding.
  • Fixed Apply writing blank files for very large cloud agent diffs. When a cloud agent changes more files than can be shown, Review Changes now warns that file contents are not available instead of saying there are no changes, and Apply is blocked instead of writing empty files.
  • Fixes to agent reliability. Follow-up messages you queue while the agent works are no longer lost, and /summarize shows an error when it fails. Agent turns no longer hang when .cursorignore rules fail to load. MCP tools that require confirmation still ask for approval when Auto-review allows the call.
  • Fixes to cloud agents. Cloud agent chats reconnect after sleep or network drops and load older turns one page at a time without stalling. Follow-ups are processed in send order and keep the model variant you picked.
  • Fixes to the agent chat transcript. Questions the agent asked in chat no longer reopen after you answer them, and large unsent drafts are no longer lost when the chat reloads.
  • Fixes to MCP sign-in. MCP servers that sign in with Microsoft Entra no longer ask for consent every time when an admin has approved the app.
  • Fixes to Git. On macOS, Cursor no longer hangs when the git on your PATH can't run, and falls back to the system or Homebrew git. Plugins load with git 2.26 and 2.27.
  • Accessibility fixes. Screen readers now announce the highlighted item and its position in the agent input's / and @ menus. Buttons in the agent chat show a focus ring when you reach them with the keyboard.
  • Admin Fixes to team and enterprise policies. Repository access rules now come only from your active team, so another team's allowlist no longer blocks your repos. Team privacy mode no longer falls back to the restrictive default during brief network errors, and Enterprise members behind unreliable proxies are no longer signed out after a temporary membership check failure.
  • Fixes to the editor. Extension hosts no longer deadlock at startup, and searches started while Cursor loads no longer fail. Ctrl+I and Ctrl+L no longer land on a chat stuck on "Loading Chat", and undo/redo no longer catches other shortcuts on AZERTY or Dvorak layouts.

Cursor mobile

Projects

  • Project chats focus on the conversation. A Project's coordinator chat shows your messages, its replies, questions, and approvals without Worked or thinking sections, and messages you send in a row are grouped together.
  • Live status while a Project works. While a Project's coordinator is running, its chat shows a status line such as the step it reported, Running commands, Messaging an agent, or Writing a reply.

Agents

  • Agent links open in the app. Tapping a cursor.com/agents link in an agent chat, including a link to a subagent, opens that agent in Cursor for iOS, and shared agent links open straight to the agent's chat instead of the inbox.
  • Earlier messages load as you scroll. Scrolling up in an agent chat loads earlier messages automatically and keeps the messages you're reading in place while they load.
  • Voice call messages are labeled. Messages sent from a voice call show a Voice call label with a waveform icon.
  • Scrub through long chats. Hold and drag the scroll indicator to move through a long conversation.

Review

  • Large diffs in Review. For very large diffs, the Review screen lists every changed file with line counts and explains when the contents exceed the viewing limit, and pull request review now loads commits past the first 50, up to 250.

Fixes

  • Projects. A Project's inbox row shows its loader while the coordinator is running, worker rows update when a worker finishes, errors, or restarts, images and media from Project workers load with a Retry option if loading gives up, and finished Projects on the Live Activity show how long the whole Project ran instead of 0m.
  • Model picker. Choosing Auto for a cloud agent follow-up switches the agent to Auto instead of keeping the previous model.
  • Large files and code blocks. Opening a very large edited file shows File too large to preview with a Copy option instead of hanging, chats with edits to large files stay responsive, and a crash while scrolling code blocks with line numbers is fixed.
  • Sending and pickers. Backgrounding the app right after sending no longer shows a false Couldn't start agent or Couldn't send message error, and the Workspace picker's Recents no longer lists repositories or environments that are no longer available.
  • Visual polish. Dragging sideways on a bottom sheet no longer moves it off-axis, large pull request numbers show without thousands separators, and overlapping avatars fully cover the ones behind them.
  • MCP sign-in failures. When signing in to a cloud MCP server fails, the sign-in card in the chat shows it as failed instead of skipped.

Cursor SDK

  • Custom tools know which session called them. The execute callback of a local.customTools entry now receives context.sessionId, the local session that invoked the tool. Subagents started with Task pass their own id, so a host can keep separate state for each child agent. TypeScript only; unset when the runtime has no session id.
  • Fixes to local agents and TypeScript types. Local agents now work on FIPS-enabled hosts, and existing local agent stores move to the new location automatically. Agents restored from the same process or VM snapshot no longer reuse request and session IDs, so concurrent runs don't collide or get stuck. The LocalSubagentInherit type declarations now resolve without an unpublished internal package, using the newly exported LocalSubagentResourceProviderFields and LocalToolExecutor types.

Cursor SDK

  • Subagents can inherit the parent's executors and tool limits. Set local.subagentInherit so that agents started by the Task tool, including nested ones, use the same custom read, write, and shell executors and reported workspace path as the parent, and the same allowed and excluded tools. Pass it on Agent.create() or override it for one send. TypeScript local agents only. If you leave it unset, subagents behave as they did before.
  • Fixed requests stalling on dropped connections in long-running agents. When an agent sits idle between runs, the SDK now closes unused HTTP/2 connections after 29 seconds. It also checks a connection that has been quiet for about a minute before reusing it. Requests no longer get written into a connection the server has already closed, where they used to hang until the stall timeout aborted them.
  • Stalled connections are retryable and bounded. When a connection stalls, the run fails with a retryable NetworkError carrying the code connection_stalled, and a streak of stall retries stops after 3 minutes instead of retrying indefinitely.
  • Fewer dependencies installed with the SDK. Installing @cursor/sdk no longer adds @connectrpc/connect-node or undici 5.x to your project's dependency tree, so installs are smaller and you no longer get version conflicts or audit warnings from those packages.

Web

  • Find merged pull requests. A repository's pull requests list has a new Merged tab with a count, and is:merged in the search box now filters to merged pull requests.
  • Request Changes reviews now block merging. A Request Changes review from someone with write access blocks the pull request from merging until it's resolved, and dismissing someone else's Request Changes review requires write access. Pending reviews you haven't submitted no longer count as approvals or blockers, and reviews on the pull request page are ordered by when they were submitted.
  • Choose repositories when requesting an app, and cancel pending requests. When you request an app install, you can pick all repositories or specific ones, with the repository from the install link preselected. A pending request shows its status, and you can cancel it from that status button. Namespace owners or team admins get an email when someone requests an install, and you get an email when your request is approved or denied. The install page also shows the app's publisher, verification, and website, a View Listing button, and the account you're signed in as.
  • IP allowlists apply to people, not your apps. When your team has an inbound IP allowlist, it now applies only to user credentials, so apps, service accounts, and agents are no longer blocked. The allowlist is also checked on Origin web pages and repository downloads, and it can hold up to 1,000 entries. Owners can read and replace the allowlist through the Origin API; see the Origin API changelog.
  • Clearer Git errors and safer pushes. When Git is denied access to an Origin repository, it now explains why, including which Cursor account was used and how to switch. Pushing a file larger than 100 MB is rejected with an error naming the file and its size. Pushes to HEAD or to ref names outside refs/ are refused instead of moving the default branch, and a malformed ref name is rejected on its own while the rest of the push lands.
  • Blame follows renames. Blame now follows file renames the way git blame does, so lines in a renamed file keep the commit that wrote them instead of pointing at the rename. Branch and tag names that look like commit hashes now resolve to the branch or tag, and commit lists and comparisons use the same merge base and ordering as Git.
  • CODEOWNERS matches GitHub more closely. CODEOWNERS patterns now follow gitignore rules, so directory and anchored patterns resolve owners the way GitHub does, and logins with underscores are accepted. Email entries resolve only to people who can review the repository. When code owners can't be determined, for example because the file is too large or too many files changed, the code owner requirement is blocked with an explanation instead of passing or failing silently.
  • Rulesets are enforced more strictly. Pull requests into a branch frozen by an active ruleset that blocks merges now show "Merging is blocked by a ruleset." unless you can bypass it. An active ruleset with a rule type Origin doesn't recognize now blocks merging instead of being skipped, and saving a rule that can't run in that kind of ruleset fails with an error. Merge when ready now waits for ruleset rules to pass even if you could bypass them.
  • Fixes to merge when ready. Merge when ready no longer merges a commit pushed after checks ran, waits for a newly linked parent pull request to merge first, and turning it off reliably stops a pending merge. A merged pull request now records the commit that actually landed it.
  • Fewer false merge conflicts. Pull requests with an edit next to a line inserted on the base branch now show as mergeable when Git would merge them cleanly, instead of reporting a conflict. When a pull request's head commit is missing, it asks you to push the branch again instead of showing an error.
  • Fixes to stacked pull requests. After a stacked pull request's base lands and it's retargeted, its diff and restack include only its own commits. Restacking keeps conflict resolutions from merges with trunk, including onto a squash-landed parent. A pull request drafted or closed at the moment its stack merged now shows as merged.
  • Fixes to diffs on large and renamed changes. Renamed files are paired the way Git pairs them, line counts match Git for final-newline and line-ending changes, and diffs for deeply nested changes now load instead of failing. On very large pull requests, Load more now fetches the next batch of files.
  • Fixes to pull request pages. Reopening a pull request whose head or base branch was deleted now tells you to restore the branch first. Permalinks now open the right file when its path has spaces or non-ASCII characters. Authors and reviewers who acted through an app show that app next to their name.

API

  • Prepare and inspect merges through the Origin API. A new call rebuilds a pull request's test merge ref against the current base and reports whether it's mergeable, conflicted, or pending, with the merge, base, and head commits. Mergeability responses also include the head commit they were computed against. See the Origin API changelog.
  • Scope check runs to a base commit. Apps can pass a base commit when posting check runs so a check counts only for pull requests with that base, and the base commit appears in check responses, the Origin SDK, and check run webhooks. Check runs posted with a failing status keep that status. See the Origin API changelog.
  • More of Origin is available through the Origin API. You can read, add, and remove reactions on pull request comments, check a user's permission on a repository, and list commits within a time range. Responses for repositories, commits, and pull requests, and pull request webhooks, now link to the page on Cursor. Ruleset bypass users now use public user IDs, which is a breaking change. See the Origin API changelog for details.

Environments

  • Environment builds are always on. Environment builds can no longer be turned off. Every environment that isn't a draft now gets automatic builds, and Cloud Agents start from the latest build. Environments that previously had builds turned off now show their active build and recurring build failures, and the Trigger build dialog no longer warns that builds are off.
  • Fix Build works on builds whose snapshot expired. Clicking Fix Build on a failed environment build whose snapshot has expired now starts the agent instead of failing. The agent boots the latest successful build and works from the failed build's logs. Launching an agent from a Dockerfile build failure also prefills the composer with a prompt to investigate and fix the build.
  • Fewer environment build failures from source control hiccups. Automatic environment rebuilds keep running after a temporary source control access failure and stop only after three such failures in a row. Builds that fail because a self-hosted source control server is down are now reported as a source control outage.
  • See which environment changes came from the API. In an environment's version history, saves made through the public API are now labeled "API".
  • Wildcard allowlist entries include the parent domain. A network allowlist entry like *.example.com now also lets Cloud Agents reach example.com itself, not just its subdomains.

Agents

  • Pinned agents follow you across devices. Pinned agents on the web agents list are now saved to your account, so your pins follow you across browsers and devices. The agents sidebar now groups agents into Today, Last 7 days, and an Older section that starts collapsed.
  • Start an agent from a closed pull request. You can now start an agent from a closed pull request whose head branch was deleted. The agent works on a new branch from the base branch, and the panel tells you so.
  • Clearer errors when an agent can't start. If an agent fails to start because you lost access to one of its repositories, the error now names that repository. When only your Other Models usage has run out, the error reads "Out of Other Models usage" and suggests switching to a named model or raising on-demand usage. When an agent's machine runs out of memory, the error suggests running memory-heavy steps on fewer files or with lower parallelism.
  • Faster agent pages on the web. Opening a Cloud Agent conversation on cursor.com no longer refetches changes that are already loading or up to date, and going back to the agents list is quicker because the list is loaded in the background.

Slack

  • Slack model picker uses the web model catalog. The Cloud Agents Slack settings model picker now builds its list from the same model catalog as the web picker and keeps your saved default selectable. The Slack model command now accepts any model you can use, instead of a fixed list.
  • Set a channel's Self-Hosted pool from Slack. Any team member can now run @Cursor pool set <name> channel or @Cursor pool unset channel in Slack. Every change to a channel's default repository or Self-Hosted pool is announced in the channel.

Fixes

  • Fixes to agent runs and follow-ups. Sending a follow-up after stopping an agent early in its first step now works instead of failing with a VM not found error. Switching between Best-of-N tabs now loads the selected candidate's conversation. Agents started on a saved environment that changed after its last build no longer fail at startup with a build version mismatch.
  • Fixes to multi-repo environments. Cloud Agents now start normally even when one of your multi-repo environments has colliding clone paths; before, that one environment made every start fail. Saving an environment whose generated name lists many repositories no longer fails.
  • Fixes to the Cloud Agents dashboard. If the Cloud Agents settings page can't load privacy settings, it now shows an error with a Try again button. Failed loads on the runs pages also offer Try again, and the runs summary reads "1 run" instead of "1 runs".

Integrations

GitHub

  • All your GitHub installations and accounts load. If you have many GitHub App installations, connecting GitHub and picking repositories now shows all of them instead of missing the ones past the first page. If one of several connected GitHub accounts has an expired token, the others still load instead of failing with "GitHub access token is no longer valid".
  • Newly created repositories. Cursor no longer denies access to a GitHub repository you can write to when GitHub briefly fails to return the repository's details, such as right after it is created.
  • Cloud Agent follow-ups on GitHub Enterprise Server. Cloud Agents started through a GitHub Enterprise Server app now receive pull request and CI updates for the team that owns the app, instead of having those webhook events rejected.
  • GitHub stays connected through GitHub incidents. When GitHub briefly rejects valid refresh tokens, Cursor keeps your GitHub connection instead of deleting it, and removes the token only if it keeps failing, so you don't have to reconnect GitHub.

GitLab

  • Clearer GitLab access errors. Starting work on a GitLab project where you lack Maintainer or Owner access now consistently says GitLab Maintainer access is required instead of asking you to reconnect, and the token cooldown error states the remaining wait in minutes or hours.
  • Connecting GitLab clears token cooldowns. After you connect GitLab, your personal and team repositories are usable right away instead of waiting out a cooldown left by earlier token errors.
  • Organization admins can manage team GitLab and Bitbucket connections. Organization admins who aren't members of a team can now load and manage that team's GitLab and Bitbucket integrations instead of getting an error.

Bitbucket

  • Replies when a comment command can't run. When @cursor can't run a command from a Bitbucket pull request comment, Cursor now replies on the pull request with the reason, such as needing a linked Bitbucket account or write access, instead of ignoring the comment.

Dashboard and admin

Cloud agents

  • Fix failed Cloud Agent environment builds in one click. When an environment build fails during install or on its Dockerfile, the environment and build pages show a Fix Build button that starts an agent to debug it, including for builds whose snapshot has expired. Builds with a deleted snapshot show a Snapshot expired chip and can't be activated, and deactivating the active build warns you when no other build can take over. Builds that fail because your self-hosted source control server is unreachable now say so. See Cloud Agent Builds.

Automations

  • Admin More control over label triggers in Approver Agents. A label-change trigger can now fire when a label is added, removed, or either, and you can limit it to one label name. These settings are now saved when you edit the agent.
  • Admin Vulnerability Scanner runs in Security Agents run history. Recent Runs on the Security Agents page now lists Vulnerability Scanner runs next to security reviewer runs.

SSO and SCIM

  • Admin More reliable SCIM directory sync. For organizations that own their SCIM directory, groups your identity provider pushes for teams linked to that organization now become synced groups instead of legacy directory groups. In directories with at least 10 users, if your identity provider briefly reports every user with no groups, directory sync no longer removes all directory group memberships.

Fixes

  • Fixes to Cloud Agent environment builds. Build logs open at the newest line and keep following new output until you scroll up. When a page of the builds table comes back empty but older builds exist, it now tells you to use Next instead of leaving you stuck, and the Trigger build dialog no longer warns that builds are off.
  • Fixes to Cloud Agents settings and runs. The Run history chart on the Cloud Agents Runs page no longer pushes the page down while it loads. Cloud agent settings refresh while the page is open, and network access entries like *.example.com now also match example.com.
  • Admin Fixes to groups and teams. On a group's Members tab, the row actions menu stays on screen on wide tables. When you create a team in an organization, the team admin list puts you first and selects you by default.
  • Fixes to the dashboard on mobile. On mobile, the navigation menu button now sits at the left of the top bar, and long GitHub Enterprise instance names truncate in settings.
  • Fixes to GitLab and Bitbucket integrations. GitLab and Bitbucket connections on the Integrations page no longer fail to load when the team selected in the dashboard is one you aren't a member of, such as a canceled team. They show as read-only instead.
  • Fixes to Bugbot settings. Old Bugbot license links open Bugbot settings in Automations.

Automations

  • GitLab support in pull request triggers and comments. The Comment on PR action accepts GitLab merge request URLs, including self-hosted GitLab. Git pull request triggers can be saved with a nested GitLab group (group/subgroup) or an organization URL as the org scope, and invalid entries show an error.
  • Clearer run failures. When a run fails because of a configuration problem, such as a missing branch, an inaccessible repository, a model blocked by your team's region, an API key rate limit, or a missing Slack channel, run history shows the specific reason and what to do. Runs whose agent ran out of memory or whose machine stopped responding say so, and a run that later succeeds shows as succeeded.
  • Fixes to automation runs. Automations with more than one schedule now run on every schedule instead of only the first. The open pull request step now opens the PR for runs owned by service accounts and for older team automations. Security Reviewer automations triggered by pull request comments or labels run again instead of being skipped, and automations with Slack access no longer stall on each step while re-listing Slack channels.
  • Fixes to team selection and PR Approver. On cursor.com, automation actions run under the team you have selected instead of your default team. PR Approver reviews no longer say reviewers were assigned when assignment was turned off or failed; they say human review is needed instead.

APIs

  • Skip the no-issues comment. POST /bugbot/review accepts an optional postSuccessComment boolean. Set it to false and a review with no new findings doesn't post the "found no new issues" comment. It defaults to true.
  • Cost for free Bugbot reviews. GET /analytics/team/bugbot-reviews now returns cost_cents for seat-based and free-trial reviews, showing their undiscounted model cost, and adds a billed field. billed is false when your team wasn't charged for the review.

Security and approval agents

  • Admin Only label. If you can't make changes to Security Agents, its pages now show an Admin Only label next to the page title.
  • Save without linking Slack. You can now save edits to a security agent without connecting your own Slack account, as long as its Send to Slack channels are unchanged.
  • More reliable security reviews. Security reviews keep results that finish right at the deadline, retry transient connection errors, and fall back to another model when the selected model is unavailable.

Models and Cursor Router

  • New model: Claude Sonnet 5.5. Claude Sonnet 5.5 is now available in Cursor.
  • New models: GLM 5.3 and GLM 5.3 Flash. GLM 5.3 and GLM 5.3 Flash from Z.ai are now available in Cursor. They're off by default; turn them on in Cursor Settings > Models.

Repository settings

  • Configure Bugbot per repository. On GitHub pull requests, Bugbot reads a .cursor/config/bugbot.yaml file to set its triggers, review effort, incremental review, PR summary, and risk score for that repository. Values in the file override team settings, and the file can lower Autofix but never turn it on. The Bugbot Settings section in the dashboard links to the docs for config files.

Autofix

  • Autofix stays off your branch in New Branch mode. When Autofix is set to New Branch, it now keeps its fixes on a separate branch and never commits to the pull request's own branch.

Reviews and checks

  • Bugbot checks no longer get stuck. A Bugbot run that is cancelled or times out now completes its check as neutral instead of leaving it in progress. When the base branch gets new commits that don't change the pull request's merge base, Bugbot still posts its review, and a cancelled run's check says why it was cancelled.
  • Clearer results for oversized and rate-limited reviews. When no file in a pull request fits Bugbot's review budget, Bugbot posts its "too large to review" comment and a neutral check instead of an error. A manual Bugbot command that hits a temporary GitHub rate limit now gets a reply asking you to retry in a few minutes. Bugbot also no longer reviews symlink targets as if they were source files.

Bitbucket

  • Better Bugbot on Bitbucket. Bugbot can now review Bitbucket Cloud pull requests from forks. When Bugbot refuses a comment command on Bitbucket, it replies with the reason. Failure comments render as markdown instead of escaped HTML, and Bugbot retries a review when Bitbucket's sign-in service is briefly unavailable instead of skipping it. The team allowlist accepts Bitbucket nicknames that contain spaces.

Dashboard

  • Clearer installation permissions in the dashboard. On your personal Bugbot dashboard, GitHub organization installations owned by a team are shown read-only, and changes are rejected with a message to switch to that team's dashboard. Personal installations you can't change without a paid plan are also read-only and say a paid subscription is required.
  • Dashboard fixes. Repository pickers and lists in Bugbot rules and learning settings show repositories as owner/name, and search matches the owner. The Set License Cap toggle reflects the saved cap after a refresh. If a Bugbot dashboard view fails to render, it shows a message with a Try again button instead of a blank page.

Billing

  • Team limits apply to token-billed Bugbot runs. On teams with token-based Bugbot billing, runs are no longer blocked by the installing user's personal spend limits; only the team's limit applies. Bugbot no longer runs as a member with the Unpaid Admin role and posts a comment explaining how to fix it. Changing Bugbot licenses on a team without Stripe billing now shows a clear error.

API

  • Skip the no-issues comment on API-triggered reviews. The Bugbot review trigger API accepts postSuccessComment. Set it to false to skip the "found no new issues" comment on that review.

Agents Window

Agents

  • Failed agents now show a clear Error status. Agents that fail now get a red Error status in the sidebar, with their own Error group and an Error option under Show > Status, on by default.
  • Usage ring in the chat status bar. When your monthly usage passes the display threshold, the chat status bar shows a usage ring with a "Usage N%" tooltip. Click it to open usage details, which now float over the conversation and close with Escape.
  • Fewer unnecessary Auto-review blocks. Auto-review no longer blocks preparatory steps such as drafting an email, staging changes, or previewing a deploy just because a block instruction forbids the later send or apply; the final committing step is still checked. Reviews also fail less often when the classifier returns an incomplete decision.
  • Pin up to 75 agents. The pinned agents limit is now 75, up from 50, so you can keep more agents at the top of the sidebar.
  • Grok 4.7 now names itself correctly. When you ask Grok 4.7 who it is, it now says it is Grok 4.7, a model trained by SpaceXAI. Before, it could call itself Composer or an earlier Grok version.
  • Voice Submit Keywords work in any language. Submit keywords in non-Latin scripts, such as Cyrillic or Chinese, now trigger auto-submit while you dictate. Accented keywords like "envíalo" also match correctly instead of being mangled.
  • Agents recover when their reasoning gets stuck in a loop. When an agent's reasoning starts repeating the same line, Cursor now stops that step and retries it once with a nudge to move on, so the run keeps going instead of stalling.
  • See where unread replies start in any agent chat. Cloud and local agent chats now mark where replies you haven't seen begin and show a New Messages pill that jumps to them, like Project chats already did.

Projects

  • Create Project keeps your work and picks the right repository. Choosing Add Models in the Create Project dialog now opens Settings without losing your project name, appearance, or model picks, and reopening the dialog restores them. The dialog no longer waits on the repository list before you can create, keeps the repository you pick from Repos or Recents instead of switching to a saved environment for the same repo, and defaults to the same environment as your New Chat draft.

Cloud agents

  • Clearer cloud agent errors with one-click recovery. Cloud agent errors now show in the tray above the input, including while a run is retrying automatically or after its session expires. When a run fails, the error offers a single next step where one applies, such as Retry, Manage Usage, or Manage Environments. Opening a terminal for a cloud agent whose machine is offline now opens a Terminal tab that explains why, and it connects automatically once the machine is reachable again.
  • Starting a Cloud Agent from a link no longer picks a random repository. Opening a Cloud Agent from an in-app prompt or a deeplink used to preselect the first repository in your list when you had no local or recent repo. Now it uses your current repo or most recent Cloud repo if you have one, and otherwise opens a new Cloud Agent with the repository picker one click away.
  • Find Bitbucket repositories beyond the first 5,000 in the project picker. In Bitbucket workspaces with more than 5,000 repositories, typing in the project picker now also searches Bitbucket, so repositories past the first 5,000 show up in the results. Browsing without a search still lists the first 5,000.

Files, changes, and pull requests

  • Review only what the agent just changed. The Changes tab now has a Last Agent Turn scope that shows only the edits from the agent's latest turn. Clicking Review or a file in the end-of-turn summary opens the Changes tab in that scope, so you can check the newest edits without digging through earlier changes.
  • Markdown previews keep your place and show agent store images. Switching away from a Markdown preview and back now restores your scroll position and cursor instead of jumping to the top. Markdown documents from an agent store also render their store images instead of broken image links.

Navigation and layout

  • More relevant command palette search results. Searching agents in the command palette now ranks chats whose transcript matches your query above chats that only loosely match the name or PR link, and typing a PR number (with or without #) still finds its chat. Searches with no matches now show "No results found" instead of suggesting a new agent, and results you can already see are selectable while other sections finish loading.
  • Screen readers can tell prompt boxes apart. Each prompt box now has its own name for screen readers, such as "Prompt", "Side chat prompt", "Subagent prompt", "Fullscreen prompt", or "Design mode prompt". In split layouts the name includes the tile number, for example "Prompt, tile 2". Menu options in the prompt box are also read as part of their list.
  • Update notifications for Linux .deb and .rpm installs. Cursor installed through apt or dnf now tells you when a new version is available instead of leaving updates turned off. Choosing Update shows the exact upgrade command with a Copy Command button, and Later snoozes that version.
  • Opening a Rust project no longer pegs the CPU. With language servers on, the Agents Window no longer starts rust-analyzer and a full cargo check just because a folder contains Rust. The server now starts when you open a Rust file, and language extensions you disabled stay off.
  • The Agents Window recovers on its own when its extension host hangs or keeps crashing. Previously, a frozen extension host could leave agent requests stuck with no error until you reloaded the window. If that extension host stops responding for 60 seconds, it is now restarted automatically and requests resume. If it keeps crashing, it keeps restarting with increasing delays instead of giving up after three crashes. The automatic restart after a hang is skipped while a debugger may be attached to the extension host.
  • Find side chats in the command palette. Searching in the command palette now finds side chats. Each one is labeled Side chat, shows its parent agent's name, and opens on its parent agent.

Customization and settings

  • MCP servers show every page of tools. MCP servers that paginate their tools, prompts, or resources now show every page, not just the first one.
  • Hooks can tie a subagent back to its conversation. The subagentStop hook now receives child_conversation_id, the ID of the subagent's own conversation, so audit and policy scripts can tie subagent activity back to its parent.
  • /subscribe cleans up after merged pull requests. A subscription to one pull request now closes on its own when that pull request merges or closes. The CI subscription on its branch closes too, unless another open pull request still uses the branch.

Fixes

  • Fixes to Projects. The project picker shows a named cloud environment's name instead of "No Repo" when its repositories are unpublished drafts. Switching between agents in a Project keeps recent conversations loaded, so going back to one no longer reloads its transcript, and with split panes the conversation in the unfocused pane stays on screen.
  • Fixes to the subagent tray. The Agents pill stays visible and shows as selected while the subagent tray is open.
  • Fixes to the composer and queued follow-ups. Queued follow-ups no longer vanish when a turn ends, and large unsent drafts are no longer cleared on reload. Typing follow-ups with keyword suggestions no longer crashes the agent panel, and scrolling while editing a previous message behaves correctly again.
  • Fixes to the chat transcript. The last messages are no longer hidden behind a pending question panel, answered questions no longer reappear, and visualization cards show an error instead of spinning forever.
  • Fixes to context summarization. After a long chat is summarized, the agent keeps your latest request instead of losing it behind an automatic reminder, and a failed summary no longer replaces your conversation history with a placeholder. Stopping an agent mid-summary now cancels promptly instead of showing a server error.
  • Fixes to cloud agents and Move to Cloud. New cloud agents keep the model you picked, cloud agents outside the current workspace no longer send false "Done" notifications, and chats that wake up in the background move up in the sidebar. Move to Cloud works for chats whose saved state lost its turns.
  • Fixes to the built-in browser. Web pages can no longer fake element picks into agent context; only real clicks select elements. Hidden browser tabs no longer keep rendering in the background, and DevTools starts only when you first open the console, so browser tabs use less CPU and memory. Browser pages and DevTools show up with their own labeled rows in Process Explorer.
  • Fixes to agent context reads. Agent context reads no longer stall when the startup .cursorignore scan fails.

Cursor IDE

Agent

  • Auto-review no longer blocks drafts and previews. When you add a block instruction like "don't send emails" or "don't deploy", Auto-review now applies it only to the call that sends or deploys. Calls that draft, stage, preview, plan, or read run instead of stopping for approval.
  • Clearer agent tool calls in chat. When the agent reads its to-do list, the chat shows the list itself. Pull request actions show their result or the error instead of being hidden, and agents blocked on approval say "Approval required" or "Approval denied" instead of "Couldn't start".
  • Agents retry when their reasoning loops. If an agent's reasoning repeats the same line over and over, Cursor stops that step and retries it once with a nudge to move on.
  • Voice submit keywords work in any language. Voice submit keywords, the spoken words that send your prompt, now save in Settings and trigger submit in non-Latin scripts like Cyrillic or Chinese. Accented letters are kept, so envíalo no longer saves as envalo.

MCP

  • More reliable MCP server sign-in and tool loading. Signing in to OAuth-protected MCP servers works with more servers, including ones that reject Cursor's redirect URLs at registration or ask you to sign in again later. Servers that paginate their tools, prompts, or resources now expose all of them instead of only the first page, and agents with many MCP tools no longer fail by going over the model's tool limit.

Hooks

  • Hooks can tell which subagent a tool call came from. When a tool runs inside a subagent, preToolUse, postToolUse, and postToolUseFailure hooks now receive parent_tool_call_id, the ID of the Task tool call that started that subagent. subagentStop hooks receive child_conversation_id, so audit scripts can link a finished subagent to its own conversation.

Browser

  • See and undo the agent's device emulation in the built-in browser. When the agent resizes a tab to a phone-sized viewport, the browser tab shows a "Device emulation · W×H" badge with a Reset button, and the emulation clears when the agent's turn ends. The browser also uses less memory because DevTools starts only when you first open the console, and web pages can no longer fake element selections or area screenshots sent to chat.

Source control

  • Git hooks run again for Source Control operations. Fixed a regression where commits, pushes, pulls, checkouts, branch creation, and clones from the Source Control view skipped repository hooks such as pre-commit, commit-msg, pre-push, post-checkout, Husky, and Git LFS. Commands that run hooks are marked (hooks) in the Git output channel. Background activity like status refreshes and autofetch still skips them.

Updates

  • Update notifications for apt and dnf installs on Linux. If you installed Cursor from a .deb or .rpm package, Cursor now tells you when a new version is available. Click Update to see the apt-get or dnf upgrade command and copy it with Copy Command, or click Later to skip that version.

Fixes

  • Fixes to agent chat and the follow-up queue. Queued follow-ups no longer vanish when a turn ends or disappear when a steer falls back to the queue. The "Worked for" duration includes tools that ran after the final reply.
  • Agent context reads no longer hang after a failed startup scan. If the first .cursorignore scan fails at startup, Cursor retries it instead of leaving agent context reads stuck for the rest of the session.
  • Fixes to cloud agents. Slash menus no longer show the previous team's managed skills after you switch accounts or teams, and cloud agents outside the current workspace no longer send false "Done" notifications.
  • Fixes to the editor and workbench. Cmd/Ctrl+Shift+B runs Run Build Task again, and unsaved backups of deleted files no longer reopen every time a window opens. Worktree fetches work on git versions older than 2.29. If the local extension host that connects Cursor to its servers hangs, Cursor restarts it so the window reconnects, unless a debugger may be attached, and it keeps restarting that host if it crashes repeatedly. Cancelled symbol and call hierarchy requests no longer leak memory.
  • Fixes to usage banners, the model picker, and chat visualizations. Models you enabled in the model picker no longer disappear when the model list refreshes. The usage-limit banner no longer stacks under a usage-limit error in the composer or shows a Request Credit button when your organization hides credit requests. Visualization cards in chat show an error instead of loading forever when their content never arrives, and chart tooltips no longer shift the surrounding chat.

Cursor SDK

  • Steer an agent while it waits on background subagents. run.steer(text) used to resolve revert_to_followup when the agent had ended its turn to wait on background subagents, so your message waited until that work finished. Now the steer runs right away as the next turn, and background results still arrive afterward. TypeScript local runs only.
  • Know when a step has finished requesting tools. onDelta now receives a tool-requests-listed update with a callCount once the model finishes listing its tool calls for a step, while those tools may still be running. Use it to tell when every tool call in a step has started, for example to group or batch a step's tool calls in your UI.
  • Fixes to cloud agent creation. For cloud agents, the first send(), which creates the agent on the server, no longer fails with an id conflict when the SDK generated the agent id: it retries with a fresh id, or continues with the agent if its own earlier create already succeeded. SDK-generated agent and run ids also no longer repeat when a process is restored from the same snapshot more than once. Ids you pin yourself still report the conflict.
  • The bridge ignores your project's .env and bunfig.toml. The standalone cursor-sdk-bridge executables no longer load .env or bunfig.toml from the working directory, which is usually your project when the SDK starts the bridge. Files in a checkout can no longer change the bridge's endpoints, tokens, or preloaded code.

Cursor mobile

Projects

  • Projects. Cursor for iOS now supports Projects. Your Projects appear in their own section of the inbox; tap one to open its chat. On iPhone, the New Project row starts a Project from a name, with an optional repository, branch, and model. If a team admin turns off Projects for the team, the Projects section and New Project are hidden.

Fixes

  • iPad chat. On iPad, the chat stays readable while the keyboard is up instead of being blurred at the bottom.

Web

  • Sign in to repositories with SSH certificates from your team's certificate authority. Team admins can add trusted SSH certificate authorities from the SSH certificate authorities section of the API Keys page, so members can clone, fetch, push, and use LFS with short-lived OpenSSH certificates instead of registering individual keys. Admins can also require certificates, which blocks registered SSH keys and API keys over HTTPS while certificates and the Origin CLI keep working. The same controls are available in the Origin API; see the Origin API changelog.
  • Build stacked pull request workflows on the Origin API. Pull requests returned by the Origin API and carried in pull request webhooks now include their stack and parent pull request, and you can set or clear a stack parent when you create or update a pull request. You can also list every pull request in a stack. See the Origin API changelog for field details and the breaking change to the old parent field.
  • Squash merges keep credit for everyone who contributed. When you squash-merge a pull request, the merge commit now adds a Co-authored-by trailer for each author of the squashed commits and for anyone those commits already listed as a co-author. Cursor's agent and bot addresses are skipped, and no one is listed twice. When a bot's pull request is squash-merged, the person who merged it and the people who approved its latest version are credited as co-authors.
  • Choose exactly where an app gets installed, and review permission reductions one by one. App install links now open a consent page for a specific namespace. If you can install the app in more than one, you pick from a "Choose where to install" list first, and the install button names the namespace you picked. When an installed app asks for fewer permissions, its page shows a "Permissions can be reduced" notice, and on the review page you can keep or remove each permission it would drop.
  • The audit log now records repository clones and fetches, and shows who edited or deleted a comment. When a person clones or fetches an Origin repository, your team's audit log now records it, along with their email and whether they used HTTPS or SSH. Background autofetches and machine actors are left out. Audit entries for pull request comment edits and deletes now name the person who made the change, not the comment's original author.
  • Fixes to merging and stacked pull requests. Pull requests with nothing left to merge are now blocked instead of landing an empty commit. Stacks with already-merged lower changes now land on the correct base, restacks list conflicted files when they stop, and a cancelled rerun no longer turns a passing check red or blocks merge when ready. Reopened pull requests also pick up commits pushed while they were closed.
  • Fixes to pull request diffs and the Files changed view. Expand all on huge files no longer freezes the tab, pull requests over the file limit let you load more files, code owner shields are back, and your pending review comments show as drafts right away.
  • Fixes to pull request pages and repository settings. Searching a repository's pull requests with an unsupported filter like is:merged now shows an empty state that explains the filter and offers Clear search, instead of silently showing the default list. Repository and namespace permission settings only list Admins and Members when they have access, and committing a suggested edit without a name and email on your profile now explains what's missing.
  • Fixes to code tours. Code tours created before a tour update now regenerate with the latest version instead of showing outdated content, while tours prepared ahead of time still appear right away. Tours no longer include stray planning notes from the model, and custom tours are left unchanged.
  • Fixes to Git operations. Rejected pushes now show the real reason next to each ref instead of a generic unpack error. Blame and file history are correct for paths with non-ASCII characters, and repositories with consecutive dots in their names now work in clone and push URLs.
  • Fixes to cancelled Bugbot checks. When a Bugbot review is cancelled on a pull request, its check now shows as cancelled instead of neutral, so a check that later becomes required can no longer pass without a finished review. Checks left on a commit the pull request has moved past still show as neutral, unless the check is required.

API

  • Apps can act on behalf of namespace members. Origin apps can now request short-lived installation user tokens that act as a member of the namespace, named by user ID or email, for API calls and HTTPS git clone, fetch, and push. Work done this way is attributed to the member "via" the app across API responses, webhooks, audit logs, and the PR page. Installations need the new User Delegation permission; see the Origin API changelog for details.
  • More of your pull request workflow is available through the Origin API and webhooks. You can now delete pull request comments, filter pull requests by head commit, see what each check run write did, and read a version's potential merge commit through the Origin API and Origin SDK. Webhooks now cover comment reactions and name who removed a label, and expired app installation tokens return a clear error. See the Origin API changelog for details.

Environments

  • Activate a draft environment build while it's still running. You can now click Activate on a draft build that is still in progress, and it becomes the active build for your environment as soon as it finishes. Activating a draft whose snapshot has already been deleted is still blocked, and no longer changes your saved environment settings.
  • Default-on environment builds work like turned-on builds. Environments with no build setting of their own, where builds are on by default, now show their Active Build and the failing-builds warning, and are rebuilt when a build is triggered, so agents in them start from a fresher build.

Agents

  • Long-running mode removed from the web. The Long-running option in the web model picker, its time-limit pill, and the team Long-running agents setting are gone. Cloud Agents started from the web always run in standard mode.
  • Coordinators can switch a worker's model. A Project coordinator can now move a worker to a different model when it sends the worker a message. The new model applies from the worker's next turn.

Self-Hosted Machines

  • Shorter self-hosted options in Slack, GitHub, and Linear. Add sh=t to a Slack, GitHub, or Linear request to run the agent on your self-hosted pool. self_hosted also accepts t/f and 1/0, and all three surfaces now take the same options. Options written inside Slack code blocks no longer start or configure an agent.
  • Take control of desktops on Self-Hosted Machines. Taking control of a Cloud Agent's desktop now reaches the machine you're viewing, including agents on Self-Hosted Machines. Older machines fall back to the previous takeover.
  • Agents on your own machines handle git and setup more carefully. Cloud Agents on Self-Hosted Machines now check the machine's git name and email before committing, and set a repository-only fallback only if none is set. They ask for credentials on the machine instead of assuming Dashboard secrets or a signed-in GitHub CLI. On My Machine, agents prefer project-local setup over system-wide installs and global config changes.
  • Slack uses a channel's default Self-Hosted pool. @Cursor launches in Slack now use the channel's default Self-Hosted pool when one is set, then the team default. If a launch is rejected, the reply names the channel default and how to change it with @Cursor pool set <name> channel.

Models

  • Grok 4.7 on the Start plan. Start plan users can now pick Grok 4.7 for Cloud Agents, fixed at medium effort and standard speed.

Reliability

  • Clearer errors when image generation fails. Image generation failures in a Cloud Agent now show a readable explanation instead of a raw status code.
  • Fixed moving some local chats to the cloud. Moving a local chat to a Cloud Agent without typing a new message no longer fails with a "Missing required fields" error when the chat's saved state has no turns. The Cloud Agent now starts idle, keeps the chat's history, and runs your next follow-up.

Fixes

  • Fixes to environments and agent starts. Starting a cloud agent no longer overwrites the environment's saved repositories, and boots the default branch when no branch is given. Failed environment builds now report the specific cause (clone, source-control access, or pod start), and agents started with agent-scoped secrets now boot from the environment's build.
  • Fixes to Self-Hosted Machines. Agents and subagents started on a self-hosted pool that accepts any repository now get picked up instead of waiting forever, and idle or deleted-run machines free up right away instead of showing as busy. Agents on Origin repositories that were detached from GitHub now receive their configured secrets.
  • Fixes to MCP connectors in Cloud Agents. Connector tools that Cloud Agents list now accept the calls made to them, including for connector accounts with custom labels. Plugin-backed MCP servers that an admin granted now serve their tools, even when the dashboard label differs from the plugin name or the agent runs as a deployment service account. The environment info tool now shows every documented environment.json field, and for subagents it reports the network egress policy that is actually enforced.
  • Fixes to agent runs and follow-ups. A follow-up sent while an agent's machine is being recreated is now saved to the queue immediately, so it no longer looks lost until the new machine is ready. An agent's link to its Origin pull request no longer disappears after a temporary lookup failure. Agents started or followed up with a specific model variant are now checked against your team's model policy using that variant, so they're no longer wrongly blocked. An agent left stuck behind a queued run that never started accepts follow-ups again after an hour instead of rejecting them.
  • Fixes to artifacts in agent pull requests. Artifact links that an agent puts inside tables or other HTML in a pull request description now render as working links instead of raw text. Pull requests no longer get public artifact links for files that couldn't be confirmed to exist, and agents with many artifacts no longer re-announce the same artifacts as new on later turns.
  • Fixes to Slack message formatting. Agent replies in Slack keep links whole, including URLs with parentheses or & and link text with brackets. Bold italic text, ~~~ code fences, task-list checkboxes, and images (now shown as links) render correctly, and text like __init__ or 2*3*4 is no longer mis-styled.
  • Fixes to the Cloud Agents page and dashboard. On cursor.com/agents, the first page of agents is now fetched on the server while the page renders, so it shows up faster on a full page load. In machine settings, the remote control toggle now explains why it is off and who can turn it on. Runs summary stat cards also reflow to fit the available width.

Integrations

GitHub

  • Repository access updates right away. After you connect, reconnect, or disconnect GitHub, changes to which repositories you can access apply immediately instead of after a cache expires.

GitLab

  • Renamed and moved projects. After a GitLab project is renamed or moved, Cursor resolves the current project instead of a stale copy.
  • Disconnecting self-hosted GitLab keeps your GitHub connection. Disconnecting a self-hosted GitLab instance on the Integrations page now disconnects that instance even when your current team can't see it, instead of disconnecting your GitHub account.

Azure DevOps

  • Bugbot repository status. The Azure DevOps Bugbot dashboard no longer shows enabled repositories as disabled.

Dashboard and admin

Cloud agents

  • Clearer Cloud Agent environment build failures. Failed environment builds now explain what went wrong, such as "Git provider isn't connected", "No access to a repository", or "GitHub IP allow list blocked access", instead of a generic error. You can activate a draft build while it is still running, and build logs now include an Ask Agent link. See Cloud Agent Builds.
  • Admin Limit Cloud Agents to selected groups. Enterprise admins can now set Restrict By to Groups in the Cloud Agents access settings and pick which groups can use Cloud Agents, instead of listing members by name.
  • Admin Set spend limits for service accounts. The service accounts list in the dashboard now shows each account's spend limit and current spend. Admins can edit an account's limit per billing period or set it to No limit.

Fixes

  • Admin Fixed the team Groups tab. The Groups tab no longer fails to load with an "invalid int 32" error for some groups.
  • Admin Fixed members and invites. The invite screen now rejects spreadsheets, PDFs, and addresses longer than 255 characters instead of creating broken invites. A newly created invite link now shows up right away instead of appearing missing, and revoking your current session from Active Sessions now signs you out instead of showing an error.
  • Admin Model access settings apply more consistently. Saving a section of the Models page no longer turns a new Cursor model in its 7-day review window on or off by accident. Only an explicit Enable now or Keep disabled choice is saved. Model blocks and release-delay settings for your team now also apply when Auto routes to that model.
  • Admin Clearer Network Access Control settings. The settings now say that the blocklist applies only to the local sandbox, list the patterns each list supports, and warn when an entry, such as an IP address with a port or a path, won't behave as expected in a sandbox list.
  • Admin Fixed GitLab for team-owned instances. On the Integrations page, GitLab repository sync and your personal connection status for a team's GitLab instance now use the team you have selected, so they load and sync correctly.

Automations

  • Fixes to Slack triggers and channel pickers. Slack-triggered automations in private channels shared across workspaces now fire for owners who are members of the channel. Slack channel pickers on large workspaces no longer miss channels after a slow load, and Refresh reloads the full channel list.
  • Fixes to automation runs. Security Reviewer automations triggered by pull request comments or CI completion run again instead of being skipped, and runs whose parent agent is no longer active fail with a clear message instead of retrying.
  • Fixes to the automation editor. Adding or editing a custom MCP server from an automation's actions no longer drops auth settings such as OAuth scopes.
  • Faster branch picker for Origin repositories. When you set a branch for an automation trigger on an Origin repository, the default branch now appears right away instead of after the full branch list loads.

APIs

  • Organization audit logs. Organization API keys can call GET /organizations/audit-logs to read audit events across the organization. Filter by time range, event type, search text, users, or a single linked team, with pagination.
  • Organization API key expiration. When you create an Organization API key, you can now set it to expire after 30, 90, or 180 days, or one year. Once a key expires, it no longer authenticates against organization APIs.
  • Look up many commits at once. The commit lookup endpoint accepts a comma-separated list of up to 100 full commit hashes. Lists with three or more hashes used to return 404; more than 100 now returns 400.

Security and approval agents

  • Smarter follow-ups on findings. Partly fixed findings are treated as addressed, and a failed follow-up check leaves the existing thread as is.
  • Overlapping findings stay visible. A new finding overlapping an acknowledged one is posted as its own comment.
  • Model fallback. Security reviews fall back to the next model when the preferred model is out of capacity.
  • Every run opens. Every run in the Security Agents run history is now clickable. A run opens its session transcript, or a run summary when no transcript is available.

Models and Cursor Router

  • New model: Grok 4.7. Grok 4.7 is now available in Cursor.
  • New model: Claude Opus 5.5. Claude Opus 5.5 is now available in Cursor.
  • Shorter Grok model names. Grok 4.5 and Grok 4.6 now appear by those names in the model picker and its tooltips, without the "Cursor" prefix.
  • Auto follows team model settings. When a team admin blocks a Cursor-served model such as Grok 4.7, or the team's release delay for new models hasn't passed yet, Auto no longer routes that team's requests to it.
  • More accurate Bugbot checks on Origin. On Origin pull requests, a Bugbot run that errors or can't post its review now fails the Bugbot check instead of reporting neutral. A run cancelled by a new push is marked cancelled, so it no longer satisfies a required Bugbot check.
  • Bitbucket-only teams can finish Bugbot setup. Teams connected only through Bitbucket no longer see a prompt to connect GitHub or GitLab during Bugbot setup. The setup steps now show the Bitbucket connection as connected.
  • Reviews continue past oversized files. When one changed file is too large for Bugbot to review, Bugbot skips that file and still reviews the others. Before, an oversized file could make Bugbot skip every file after it.

Agents Window

Enterprise teams get Projects from 3.21.9 unless opted in.

Agents

  • Team rules with file patterns now apply only to matching files. Fixed team rules scoped to a file pattern, such as *.py, being added to every agent conversation. They now apply when the agent reads a matching file or you add one to context, so unrelated chats no longer spend tokens on them. Team rules without a file pattern still apply to every conversation.
  • Fixed getting stuck when a usage limit offers Switch to Auto. When a usage limit offers a model switch, the limit banner now shows Switch to Auto next to the upgrade button, and the prompt stays editable. Choosing Switch to Auto clears the banner and resends your message with Auto so the conversation can continue.
  • Cloud subagents waiting on you now stand out. When a cloud subagent stops to ask you a question or needs you to sign in to an MCP server, the Agents pill shows a yellow status dot and the subagent's row in the tray is marked as needing attention, so you can unblock it instead of waiting on a stalled run.
  • Video review subagents no longer review videos they weren't given. When the agent asks a subagent to review a video, the video file now has to be attached. A file path that only appears in the prompt no longer counts. If no video is attached, the subagent says so in plain text instead of making up a review.

Chat and composer

  • A Save check mark while editing a queued message or goal. When you edit a queued message or a goal, the send button becomes a Save check mark.
  • Smoother scrolling while an agent streams. While the chat follows a streaming reply, it now glides to new text and tool cards instead of jumping.
  • The steering prompt opens Settings. The Steer New Messages? banner above the queue now opens Settings at New Messages instead of changing the setting for you. The first time a message steers a running agent, a one-time notice says so and links to the setting.

Projects

  • Create Project picks a smarter default and explains failures. The Create Project dialog now preselects the cloud repo or environment of the agent you're viewing, or of your New Chat draft, including multi-repo environments. When creating a project fails, the dialog shows the reason instead of a generic retry message, and if your plan blocks it, the notification includes an Upgrade button.
  • Project workers get renamed when reassigned. When a Project's coordinator gives a worker a very different task, it can now rename that worker so its name in the sidebar matches the new work.

Cloud agents

  • See which self-hosted worker runs your agent. Agents running on a self-hosted pool show the worker's name and its pool in the chat status bar.
  • Find any cloud environment by name from the project picker. Typing in the project picker searches all of your cloud environments, including ones beyond the first page, and highlights the matching text. The Run on menu shows Cloud with a loading spinner while access is still being checked instead of hiding it.
  • Attach PDFs, code, archives, and other files to cloud agents. Cloud agent prompts now accept any non-image file up to 10 MB, from drag and drop or the + menu, so the agent can read it in its environment. Videos and empty files are not supported yet, and you'll see a message when a file can't be attached.

Files, changes, and pull requests

  • Pull requests for Origin repositories open in the Agents Window. Clicking a github.com pull request link for a repository tracked on Origin now opens the PR tab in the Agents Window instead of your browser, when PR Link Destination is set to open inside Cursor. Bugbot and Automations comments on Origin pull requests now show native Fix with Agent and Add to chat actions instead of raw image buttons.
  • Browser tabs you open stay with your workspace. Browser tabs you open yourself in the Agents Window now persist with the workspace and stay visible as you switch between agents, like your own terminals. Browsers an agent opens stay tied to that agent. Each workspace also keeps its own cookies, logins, and storage, and signing out of Cursor clears browser data for every workspace.
  • Syntax highlighting now matches your editor theme. Code in diffs, PR review cards, and the plan tab is highlighted with the same grammars and theme colors as the editor. It updates when you change or import a theme, and no longer drops colors like numeric literals.
  • Markdown previews render math and remember where you were. Display math wrapped in $$ now renders as formatted equations in the Markdown preview. Previews also keep their scroll position and cursor per file when you switch tabs and come back.
  • Terminals are easier to reach from more places. The Terminals pill and platter for an agent's terminals and background shells now also appear in the fullscreen editor's agent overlay. Clicking a headless background shell now opens its output in a tab, Toggle Terminal hides the panel again when a terminal is already open, and the Terminals pill on Project agents opens and closes the terminals tray again.
  • Fresher pull request status. Cloud agent pull requests now refresh CI checks, reviewers, and mergeability every few minutes, so a stale Debug CI Failure pill clears once checks go green.
  • Fixed missing Git remotes when your Git config uses includes. Remotes defined in files pulled in with include or includeIf in your Git config are now detected. Remote URLs that contain spaces, and config sections with comments, are also read correctly.

Navigation and layout

  • Pinned cloud agents sync across devices. Cloud agents you pin are now saved to your account for everyone, so they stay pinned across windows and devices. The sidebar's Pinned section also shows them right away at startup instead of waiting for the agent list to load.
  • Better agent search ranking in the Cmd+K palette. When you search agents with Cmd+K, agents whose conversation contains your search now rank above agents that only loosely match by name or PR link. Typing a pull request number still finds its agent.
  • Open IDE opens the folder you picked. After running cursor . or choosing a folder with Open Folder in the Agents Window, Open IDE now opens that folder in the editor instead of an empty window, even before you start an agent.

Customization and settings

  • Fixes to MCP approvals and sign-in. The MCP tool approval prompt no longer offers "Always Allow" when your team's settings or a permissions file manage the allowlist, where the choice had no lasting effect. MCP servers whose saved sign-in was rejected, such as X, now show as needing authentication with a login option instead of failing to load. MCP sign-in cards in cloud agents are no longer dismissed early, and cards you already answered no longer reappear.
  • Close the Automations panel with Esc or Cmd+W. The Automations panel now closes on Esc or Cmd+W (Ctrl+W on Windows and Linux), just like Settings and Customize. Before, Esc did nothing there, and Cmd+W could prompt you to close the whole window. Esc still goes back one step when you're inside an automation, and it won't close the panel while you're typing in a field.

Fixes

  • Fixes to agent reliability on long runs. Cloud agents that keep repeating the same tool call now get the same nudge to break the loop that local agents get. Long turns with many tool calls no longer get stuck summarizing context over and over without freeing any space. File edits no longer fail when a model sends its patch in a different format.
  • Fixes to cloud agent workspaces and slash menus. Cloud agent workspaces that lose their connection are repaired in place instead of needing a window reload. After you switch accounts or teams, cloud agent slash menus no longer show the previous team's skills.
  • Fixes to the sidebar. Clicking + on a repository section starts the new agent on exactly that section's repositories. Project rows show the unread dot even while child agents are still working and no longer show an expand chevron when the Projects section is collapsed, and titlebar icons no longer shift when the sidebar collapses.
  • Fixes to workspace search and SSH workspaces. Workspace search no longer hangs on cloud and remote workspaces. It shows a message while the workspace connects or reconnects, and ends with a clear message if the remote host stops responding. Adding an SSH workspace with several remote folders is also faster.
  • Fixes to the model picker. Saved MAX-only model variants, like 1M context, now reset only when you turn MAX off, not when you use the picker in a non-MAX chat. When your team lifts a "Restrict to Auto" model policy, the picker no longer keeps showing the restriction notice or hiding the search bar.
  • Fixes to the browser. Open in External Browser on a local HTML file opens your system browser, and when an admin blocks browser automation your saved browser mode is no longer reset to Off.
  • /subscribe works with Origin repositories. Cloud agents in a repository hosted on Cursor Origin now wait for its CI results and pull request activity with the Origin subscription tools instead of polling. The agent picks the tool that matches the pull request link, so an Origin link no longer goes to a GitHub tool and fails.
  • Fixes to the agent panel. The agent panel no longer gets replaced by a "Something went wrong" screen when a chat that is still open is unloaded in the background. The chat stays on screen.

Cursor IDE

Browser

  • Separate browser sessions per workspace. The built-in browser now keeps cookies, logins, and storage separate for each workspace, so sessions from one project no longer carry into another. See session persistence. Signing out of Cursor clears browser data for every workspace. Element selection and area screenshots respond only to your own clicks, so a web page can't inject fake selections into chat.

Agent

  • Agent responses start faster. Cursor prepares the agent in the background when you focus a chat and again when you send a message, so the first response arrives sooner.
  • Admin Team rules with file patterns apply only to matching files. A team rule scoped to a pattern such as *.py used to be added to every agent conversation. Now the agent picks it up only when it reads a matching file or you add one to context. Team rules without a pattern still always apply. See how Team Rules are applied.
  • The @ menu no longer shows non-working Docs options. The Docs category and Add New Custom Docs option had already stopped adding documentation to agent context, so they're gone from the @ mentions menu. Docs chips in older conversations still display.
  • The steering prompt opens Settings. The Steer New Messages? banner above the queue now opens Settings at New Messages instead of changing the setting for you.

Editor

  • Math and better code highlighting in the Markdown editor. The Markdown editor renders $$ display math blocks as formatted equations. Code blocks in the Markdown and plan editors get the same syntax highlighting and theme colors as the code editor, and update when you switch themes.

Fixes

  • Removed lines stay removed in agent edit diffs. In the diff shown for an agent edit, a removed line that starts with "--" stays a removed line.
  • Fixes to editing previous messages. Editing and resubmitting an earlier message reliably sends your edited text and refreshes checkpoints. While editing in place, Cmd+/ opens the model picker for that message, Escape closes an open picker without cancelling the edit, Shift+Tab works, and the Cmd+Enter-to-submit setting is respected.
  • Fixes to Cloud Agents. Cloud agent workspaces no longer show the folder trust prompt or a file-watcher warning you can't fix.
  • Fixes to agent reliability. Long turns no longer get stuck compacting context on every step, and agent edits no longer fail when the model wraps a patch in JSON. Chats stay attached to their folder after Add Folder to Workspace, and unread badges on agent chats clear when you focus them.
  • Fixes to the model picker. Saved model variants that need Max Mode, like 1M context, no longer reset when you use the model picker in a chat without Max Mode; they revert only when you turn Max Mode off. When a team lifts its "Restrict to Auto" model policy, the picker drops the old restriction notice and brings back the search bar.
  • Fixes to MCP servers. The MCP approval prompt no longer offers "Always Allow" when your team or a permissions file controls the MCP allowlist. MCP servers whose saved token was revoked ask you to log in again instead of showing a connection error. MCP sign-in cards in cloud agents are no longer dismissed early, and cards you already answered no longer reappear.
  • Fixes to git and the editor. Git remotes defined through include or includeIf in your git config are detected, and the status bar blame always matches the active editor. The extension host no longer fails to start when an environment variable name begins with a digit, and backups for deleted or unreadable files no longer reopen on every window open. In the Review Changes pane, an older refresh no longer overwrites the newer Issues Found list.
  • Search in remote workspaces no longer hangs. When search isn't available yet, it now times out and shows an error instead of waiting forever.
  • Fixes to agent visualizations. Visualization cards from /visualize now render in the IDE. Shaded areas in charts no longer render solid black, and the visualization loading row lines up with other tool calls.

Cursor mobile

Composer

  • Attach any file type from Files. When you add context in the composer, the Files picker now lets you pick any file, not just a fixed list of types. Images attach as images, and other files such as archives, audio, or source code attach as documents. Videos aren't supported and show a message instead.
  • See your words while dictating. When you dictate in the composer, a live caption shows your words as you speak, with the part still being recognized in gray italics.

Agents

  • Preview cards in agent chats. When a cloud agent shares its running app, Cursor for iOS shows a preview card with a screenshot, title, and description. Tap Preview to open the agent's desktop view on cursor.com.
  • MCP sign-in requests. A cloud agent waiting for you to sign in to an MCP server shows as Needs Attention in the inbox and on the Live Activity until you tap Skip or Authenticate. The sign-in card then says the server is authenticated or that authentication was skipped.

Agent list

  • Collapse Pinned and Drafts on iPhone. Tap the Pinned or Drafts header in the agent list to collapse or expand that section. The setting carries over to the iPad sidebar.

Review

  • Review Origin pull requests. On Cursor Origin pull requests, a Review button opens a sheet where you can Comment, Approve, or Request Changes with a message.

Fixes

  • Agent chats. Links in agent messages open as soon as you tap them, and a new agent's prompt stays pinned to the top instead of jumping when the reply starts. Partly loaded finished conversations reload without duplicating messages, and the loading animation's dots are visible again.
  • Pull requests and repository names. Tapping Mark Ready on a draft pull request opened from a link now marks it ready for review. Repositories hosted on Cursor Origin or with URLs ending in .git show as owner/repo.
  • Composer and keyboard. Renaming an agent no longer pushes the composer bar up with the alert's keyboard. The composer placeholder changes smoothly between new-agent and follow-up text, and the quick-action pills above the composer fade in place, with a plain fade when Reduce Motion is on.
  • Copy and visual polish. The profile sheet shows chart titles and gridlines while stats load. In-progress labels read "Working" and "Thinking", and the Customize sheet sections are renamed Show and Agent Properties.

Web

  • Clearer repository access, and access is removed when people or groups go away. The Permissions tab in codebase and repository settings is now called Access, and each group there shows whether it's a Team Group, Org Group, or Synced Group from your identity provider, with its member count. When someone leaves a team or organization, they lose the repository access they had through its groups. When an organization group is deleted, the repository and namespace access it granted is removed automatically.
  • Emails when team or group changes affect your repo access. When you join or leave a team or group, Origin now emails you once per change, listing the repositories and codebases you gained or lost access to. Each resource in access emails links to its page in Origin.
  • Redesigned Apps pages with a Marketplace and clearer install screens. The Apps page in your codebase settings now opens to a Marketplace for browsing apps, with a separate Manage view for installed apps. The install screen shows whether an app is Verified by Cursor, meaning it's listed in the Marketplace, or Not verified by Cursor. It also has a simpler choice of repositories. App pages can list the publisher's public Cursor plugins, and every viewer of a repository's Apps settings can open installed apps' pages.
  • Faster pull request, repository list pages. Pull request pages now show the title, status, branches, Auto-merge badge, and existing reviewers on first paint instead of placeholders. Repository lists and a repository's Pull requests list render their first page with the page.
  • Build pull request searches with suggestions. The search box on the pull requests list now suggests qualifiers like author:, label:, assignee:, review:, is:, and sort: as you type, along with matching values such as repository collaborators and labels. Recognized qualifiers are highlighted in the box so you can see which filters apply.
  • Submodule updates show up properly in pull request diffs. When a pull request adds, updates, or removes a submodule, the Files changed tab now shows a submodule card with the old and new commits instead of an error or an empty file. When both commits come from the same repository on GitHub or Origin, the card links to a comparison between them.
  • Repository downloads start right away and extract into one folder. Downloading a repository archive from the codebase page now returns the file directly, even when it has not been prepared yet, so you no longer need to wait and retry. Archives extract into a single owner-repo-shortsha/ folder and download with that name, matching the layout other Git hosts use. See the Origin API changelog for details.
  • Clearer names for Bugbot and Security Review checks. Bugbot checks on pull requests now appear as "Cursor / Cursor Bugbot" and "Cursor / Cursor Bugbot Autofix" under one Cursor group, instead of repeating the same name twice. Security Review checks now appear under "Cursor Security Agent" with the automation name as the check title, instead of an internal account name.
  • Create a repository from the Codebase welcome screen. If you can create repositories, have none yet, and haven't dismissed the Codebase welcome screen, it now has a Create Repo button under Sync from GitHub that opens the New Repo dialog.
  • Safer pushes, app installs, and rebases. Pushing a branch or tag whose name isn't valid UTF-8 now rejects only that ref with a clear error, while the rest of the push lands. Creating or installing an app is now refused for namespaces that can't write to Origin, matching repository creation. Commits Origin replays, for example when rebasing a pull request, now keep their jj change-id so jj users keep change identity.
  • Fixes to stacked pull requests and retargeting. Retargeting a pull request to the base it already has, or to a merged ancestor's branch, no longer detaches it from its stack, and merged stack members now show the branch they landed on as their base. Restacking works on branches that contain merge commits, CODEOWNERS review requests after a retarget match the rules the merge check enforces, and reopened pull requests run CI against a fresh merge with the current base.
  • Fixes to pull request diffs and the API. Diff timeouts now show a timeout error instead of a generic error, the HTTP API :syncMirror endpoint works again, and browsing a repository with no commits reports an empty repository error.
  • Search commits by SHA. A repository's Commits page now has a search box. Type at least five characters of a commit SHA to preview matching commits, or press Enter to list the commits that start with that SHA.

API

  • Search file contents through the Origin API. A new Grep Contents operation in the Origin API searches the files in a repository at any ref and returns the matching lines. You can use a regular expression or exact text, ask for surrounding context lines, and narrow the search with include and exclude globs. See the Origin API changelog for details.
  • Check whether a pull request can merge from the API. A new preview endpoint in the Origin API tells you whether a pull request is mergeable and, if not, what is blocking it: required checks, approvals, code owner approvals, merge conflicts, or the shape of its stack. You can pass an expected head commit so the answer fails fast if the pull request has moved. See the Origin API changelog for details.
  • Sort pull requests by recent activity and list only merged ones. When you list pull requests through the Origin API, you can now sort them by last update instead of creation time and filter to merged pull requests only. Sorting by update time also works in the Origin MCP server's pull request listing tool. See the Origin API changelog for details.
  • Delete branches through the Origin API. You can now delete a branch from a script or integration with the Origin API, without pushing a deletion. The default branch and branches in mirrored repositories can't be deleted this way, and open pull requests from the deleted branch are closed. See the Origin API changelog for details.
  • Add repositories to an existing app installation through the Origin API. Namespace admins can now add repositories to an app installation through the Origin API without going through the install flow again. The call only adds repositories and never changes the app's permissions. If the app isn't installed yet, the error now explains that an admin must finish the first install in the browser and includes the link to open. Listing an installation's repositories now also accepts a case-insensitive filter on repository name or owner/repo. See the Origin API changelog for details.
  • Webhooks for pull request labels. Origin apps can subscribe to pull_request.label.added and pull_request.label.removed events, so your automations can react when labels change on a pull request. Each event includes the pull request and the label, and label-added events also include who added it. See the Origin API changelog for payload details.
  • Email alerts when webhook deliveries pause. If your Origin app's webhook deliveries are paused automatically after repeated failures, namespace admins and the people who created or last edited the app now get an email with a failure summary, recovery steps, and a link to manage deliveries. Receivers now have up to 10 seconds to respond before a delivery times out, up from 5. Check run created and completed payloads no longer include a top-level actor; use check_run.actor instead. See the Origin API changelog.
  • Scope changes for apps and repository creation. Reading app details through the Origin API now requires the namespace apps read scope, and creating or mirroring repositories requires the "Create repositories" scope. See the Origin API changelog for details.

Agents

  • Follow-ups to agents archived elsewhere now go through. If you archive an agent on another client, such as the desktop app, and then send a follow-up from a tab you already had open on the web, Cloud Agents now unarchives the agent and delivers your message instead of failing. This works from the agent chat, resubmitted messages, side chats, and the review agents sidebar on pull request pages. Deleted agents stay closed.
  • Plan and limit dialogs appear where you start the agent. When a Cloud Agent can't start because of your plan, usage pricing, model choice, or a protected scope, the dialog explaining why now opens on the page you started from, for pages within the agents area of cursor.com, instead of later on the agents page.
  • Draft pull requests stay in draft. Fixed an issue where Cloud Agents marked their draft pull requests ready for review on their own, even when your repository rules said to keep them as drafts. Agents now follow your instructions and repository guidance, and still mark a PR ready when you ask.
  • Fixed agents reading the default branch when started on another branch or commit. When you start a Cloud Agent on a specific branch or commit, it now waits until that ref is checked out before running any tools, so it no longer reads files from or describes the default branch first.
  • Cloud Agents keep their environment and files when it is moved to another machine. Previously, if an agent's environment was moved while you sent a follow-up, the agent could report that the environment was unreachable and continue in a fresh one, losing your files. Now the agent shows "Reconnecting to your environment. This can take a few minutes.", retries on its own, and continues in the same environment with every file in place. If reconnecting takes too long, it asks you to send your message again instead of starting over.
  • Pin up to 75 agents. The agents list on cursor.com now lets you pin up to 75 agents, up from 50.
  • Blocked models are now rejected as soon as you send a follow-up. If you switch a Cloud Agent follow-up to a model your team admin has blocked, you now get a Model Blocked error right away. Before, the follow-up was queued and then failed mid-run or quietly ran on a different model.
  • Fixes to page layout on the Agents web app. Opening a conversation, repo, or pull request no longer makes the page jump when the sidebar collapses, and the sidebar loads already collapsed if you keep it that way, or when you open a conversation in a narrow window. The agents list no longer shifts sideways when its scrollbar appears, or moves down when promotions, the Cursor Learn banner, or the connect source control prompt show up after the page loads.
  • Fixes to Cloud Agent runs and follow-ups. Queued follow-ups, repo skills, and slash commands now load in side chats, and repo command pills from the composer pass their contents to the agent. PR creation failures from GitHub rate limits now say how long to wait.
  • Fewer failed starts and follow-ups. When a Cloud Agent can't get a worker to start its session, it now retries on its own; if it still can't, it says "Cloud Agent could not get a worker to start your session; please retry in a few minutes." Follow-up turns no longer pick up another account's personal rules or skills from the workspace. The slash command menu of a failed or archived agent stops waiting for commands that will never load, and folders deleted from an agent's Context no longer reappear with leftover files.
  • Faster starts on a named branch. Cloud Agents started on an explicitly named branch, whether the default branch or another one, can now use the faster startup path instead of falling back to the older start.

Projects

  • Team admins' Projects setting now applies to Cloud Agents. When a team admin turns off Projects, the team's Project agents no longer show up in Cloud Agent lists, and starting a new Project fails with "Projects are disabled for your team." Personal accounts, and teams that haven't changed the setting, keep Projects on.
  • Fixes to Projects coordinator and worker agents. Coordinators can now start workers after a Project's repository is published under a new name, and PR label edits, environment drafts, and shared links use the published name too. Force-submitting a message to a worker no longer leaves the coordinator's call to it stuck as running. When a worker can't be created, the coordinator gets the real reason, such as hitting the nesting limit, instead of retrying a generic error.
  • Cloud workers share the Project's Context. Cloud workers that a Project coordinator creates now use the coordinator's Context as their own, so files they save there are visible across the Project. Moving a worker out of the Project gives it its own Context again.

Environments

  • Fixes to environment builds. Builds stuck in progress no longer block scheduled builds, and stuck builds are marked failed after 9 hours instead of 24. The environments list also loads for large teams instead of timing out. Team environment builds that failed because the team's build account couldn't access a repository now rebuild with the access of the person who started the agent.

Self-Hosted Machines

  • Self-hosted pools now respect repository access. Starting a cloud agent on a team's Self-Hosted Machines pool now fails with an access error if you can't read the repository. Pool listings only show the repositories you have access to, and API keys limited to certain repositories can only use pools for those repositories.
  • Self-hosted agents without a repository can set one up. Agents on Self-Hosted Machines that start without a repository now follow your worker's rules and credentials to find or clone the repository they need, instead of saying they have no repository access. If no rule explains how to get the repository, the agent says it isn't configured rather than guessing a clone URL.
  • Fixes to self-hosted and private workers. Agents on pools whose machine is still booting wait for it instead of failing, dead worker connections reconnect or reassign instead of failing with a closed-transport error, and agents on Windows workers no longer break the clients that display them. Subagents sharing a parent's worker no longer orphan the parent's workspace.

Integrations

  • More reliable self-hosted routing from Slack, GitHub, and Linear. Mentioning the word self_hosted in a request no longer sends the agent to a Self-Hosted Machine by accident. Opt in explicitly with self_hosted=true, which now also works inside inline code and with trailing punctuation. In GitHub and Linear, options inside fenced code blocks are ignored. In Slack, pool=<name> for a self-hosted pool with no repository now starts the agent on that pool instead of failing to match the team's default repository.
  • Jira launches use the right repo host. Cloud Agents launched from Jira with a default repo set as owner/repo now run against that repo on your connected GitLab or Azure DevOps host, instead of assuming github.com.

MCP

  • Cloud agents with many MCP tools no longer fail on model tool limits. When your MCP servers expose more tools than a model accepts, cloud agents now cap the MCP tools they send, keeping browser and custom tools first, instead of failing the request.
  • Clear error for empty Cursor-hosted repositories. Starting a cloud agent or environment build on an empty repository hosted on Cursor now says the repository is empty and asks you to add an initial commit, instead of showing a generic failure.

Dashboard

  • Fixes to Cloud Agents settings and machines pages. The My machines list no longer shows the previous team's machines after you switch teams, and its page is kept in the URL so links and back navigation land on the right page. Closing the New environment dialog in any way, including browser back, now clears the draft instead of leaving stale repositories, name, and scope. Settings sections such as Preferences and Security have shareable anchor links, and counts show thousands separators.

Integrations

Slack

  • Full bot alerts in automations. Automations triggered from Slack now read the full content of bot alerts that use attachment blocks, instead of only their one-line fallback text.

GitLab

  • Large GitLab groups and repositories. The GitLab repository picker no longer times out on large organizations, and resolving a branch or tag works on repositories with thousands of branches.

Azure DevOps

  • Reconnect prompt for expired Azure DevOps access. When Microsoft Entra rejects your Azure DevOps refresh token, Cursor stops retrying and the dashboard shows the connect card so you can reconnect.
  • Large Azure DevOps access tokens. The dashboard and Cloud Agent starts no longer fail with a server error when your Azure DevOps access token is larger than 4 KB.
  • Per-organization errors in the Azure DevOps repository picker. When one Azure DevOps organization refuses to list its projects, the repository picker shows an access error on that organization instead of failing the whole picker.

Dashboard and admin

Members and groups

  • Admin Groups pushed from your identity provider now appear as team groups. When your identity provider pushes a directory group through SCIM, Cursor creates a matching team group in the dashboard and keeps its members in sync. Removing a user from the group in your identity provider removes them from the team group, and deleting the group retires it. Groups that synced before this change keep working as they did.
  • Admin Removed members are now signed out everywhere. When you remove a member from a team that isn't connected to SSO, Cursor ends all of their active sessions so they lose access right away. Sessions stay active only if the member still belongs to another team in the same organization. Your audit log stream also gets a new credentials revoked event that records what happened to the member's sessions, repository access, and cloud agent git tokens.
  • Choose your home team in an organization. If you're a direct member of two or more teams in your organization that have billing set up, at least one seat, and a subscription that isn't canceled, pick a home team under My Settings > More while a team in that organization is selected in the dashboard. Cursor uses that team for your account when no team is specified, or you can set it to None.
  • Admin See a group's Codebase access on a new Access tab. Team and organization group pages have an Access tab that shows the group's access to your team's Codebase and any repository-specific access, with Manage links to change it.
  • Admin Organization admins appear in team member lists. Team member lists now include admins of the linked organization, labeled Org Admin.

Workspace settings

  • Admin Turn Projects on or off for your team. Admins can use the Enable Projects switch in Team Settings, under Chat, to control whether team members can use Projects in Cursor. When it's off, Projects is hidden for everyone on the team.
  • Admin Data exports for every team. The Data Exports tab is no longer limited to Enterprise teams. Any team member with permission to read exports can open it and download archives Cursor has prepared and released for the team, including teams whose subscription has ended. Export links open straight to the right team.
  • Admin Team MCP servers on members' machines (since removed from the dashboard). This week added a Run team MCP servers on members' machines switch to the Team MCP Servers section of the Plugins & MCPs tab. The Plugins & MCPs tab no longer shows that switch. Teams and members can also configure up to 1,000 MCP servers each.
  • Admin Connect more than one Jira site to a team. Admins with a connected Jira site now see a Connect another site action on the Jira page in the dashboard. It opens the Atlassian Marketplace to install Cursor on the new site, and the site list refreshes when you come back.
  • Clearer way back to Agents from settings. The settings sidebar header now shows a Back to Agents link next to the Cursor mark.

Usage and cloud agents

  • Admin Security Reviewer spend is easier to track on the Usage page. Security Reviewer usage is now attributed to Security Reviewer, and you can filter usage events by it. When you filter by Security Reviewer, each review run shows as one row, including the reviewer agents it spawns, instead of being split across many rows.
  • See exactly what a cloud agent environment build used. Environment build pages now have a Configuration section that shows the build's source, install and start scripts, terminals, and environment.json. Build logs sit in a collapsible Build output panel with an Ask an agent about these logs button, and they keep streaming past one minute without disconnecting or repeating lines.

Fixes

  • Dashboard pages no longer jump while they load. Overview, Usage, Billing, Cloud Agents, and Integrations pages reserve space for banners, usage tables, and invoices, and the dashboard no longer shifts sideways on first load.
  • Admin Fixed MCP server and secrets settings. The Choose Tools dialog stays on screen with its header and Use Selected Tools button visible, scrolls long tool lists, and trims long tool descriptions to three lines with the full text on hover or click. Editing an MCP server no longer drops extra auth settings such as scopes. The update warning for a new secret now appears only when a secret with that name exists in the same scope.
  • Admin MCP restrictions on team groups now apply. MCP settings saved on a team group are now enforced for the group's members. Editing MCP settings on a SCIM directory group linked to a team group saves them to that team group without dropping its existing restrictions.

Automations

  • Enable Cursor-built agents from a card grid. The From Cursor section of the Automations page shows each agent as a card with an enabled or disabled badge and an Enable or Manage button.
  • Clearer delete confirmation. Deleting an automation from the list or its settings page opens a dialog that names the automation and warns that deletion can't be undone, replacing the browser's confirm popup.
  • Fixes for Security Reviewer. Security Reviewer keeps syncing when a trigger repository is deleted or removed from the GitHub App, covers the repositories in an organization-wide trigger, and retries Origin, GitLab, and Bitbucket repositories that failed to connect.
  • Automations on Origin repositories. Security Reviewer now triggers on pull requests in team Origin repositories. Origin check runs from automations are grouped under the product name, such as Cursor Automation, and each run is named after its automation.

APIs

  • Find groups by ID or name. Organization group objects now include a publicId, and List Organization Groups accepts a name query parameter that returns the group with that exact name.
  • Filter organization members. GET /organizations/members now accepts email, userId, search, organizationRole, and teamId query filters. email and userId each take up to 100 comma-separated values.
  • Team directory groups. The Admin API adds /teams/directory-groups endpoints to list, create, read, update, and delete a team's directory groups, including each group's name and monthly spending limit, and to add or remove members by user ID.

Security and approval agents

  • MCP actions after a review. When a review finishes, Security Reviewer carries out the actions your custom instructions ask for, using the agent's configured MCP servers.
  • Redesigned run transcripts. Opening a Security Reviewer run that ran on Cursor-managed hosting shows a transcript styled like Cloud Agents.

Models and Cursor Router

  • 1M context kept on Cloud Agent follow-ups. A Cloud Agent started on the 1M-context variant of GPT-5.6 Terra now shows 1M in the web follow-up model picker, and follow-ups run at 1M instead of 272K.
  • Clearer Bugbot check names on Origin. On Origin pull requests, Bugbot's checks are now named "Cursor / Cursor Bugbot" and "Cursor / Cursor Bugbot Autofix". Required-check rules that already list Bugbot keep matching.

Web

  • Stacked pull requests merge as one commit, and your commit text lands. Merging a pull request in a stack now lands it together with the open pull requests below it as a single commit. When you merge a single pull request, the commit title and message you edit in the merge popover now land exactly as written.
  • Pause, resume, and test app webhook deliveries. Your app's settings page now has a Webhook Deliveries section. Once the app has a webhook URL, the section shows whether deliveries are active or paused, and has Pause, Resume, and Send Test Delivery buttons. Events that queue up while deliveries are paused are dropped, not sent late when you resume. Origin may also pause deliveries on its own, but only after a receiver has failed every delivery for 72 hours, with at least 20 failed deliveries in that time.
  • Codebase settings has a sidebar and linkable sections. Codebase settings now uses a sidebar instead of tabs, with a Back to Codebase link at the top. Each section, such as Permissions and Apps, has its own URL you can bookmark or share, and older settings links redirect to the right section.
  • Download a copy of your repository. Repository settings now have a Data section on the Advanced tab with a Download repository button that saves a tarball of the latest default branch. Origin API clients can request the same archive for any ref; see the Origin API changelog.
  • See the pull requests behind a commit. When a commit was pushed as the head of an Origin pull request, or is the merge commit Origin created when landing pull requests, its commit details now include a References section listing those pull requests' branches, with links to each branch and its pull request. Each pull request shows whether it is open, draft, merged, or closed, and long lists collapse behind More and Less. References appear on repositories hosted on Origin, not on repositories mirrored from GitHub.
  • Easier-to-scan pull request lists. Pull request lists now show larger titles that wrap instead of truncating, with a line under each title showing the pull request number, its author, and how long ago it was opened. Rows highlight on hover and match the style of the commits table.
  • Image previews in pull request changes. Added or changed PNG, JPEG, GIF, and WebP files in a pull request's Changes tab now show a preview instead of an empty binary file card. Transparent images sit on a checkerboard background, so light icons stay visible in light theme and dark icons in dark theme.
  • Safer pushes to Origin repositories. Origin now refuses a git push that deletes a repository's default branch, with the error "refusing to delete the current branch", matching Git and GitHub. Other branches can still be deleted. Pull request refs under refs/pull/ and refs/changes/ no longer appear when you push and can't be overwritten by a push.
  • Pull request bases must be existing branches. Creating or retargeting a pull request now rejects a base that is a commit SHA, a tag, or a branch that doesn't exist, with an error saying the base must name an existing branch. Before, these bases were accepted but left the pull request unable to get a merge ref for CI or to be merged.
  • Merge when ready no longer gets stuck on pull requests that can never merge. If a queued merge fails for a permanent reason, such as missing permissions or a pull request that no longer exists, merge when ready now turns itself off instead of retrying indefinitely. Merge conflicts and a changed head still leave it on.
  • Re-run checks on pull requests. Completed checks whose app allows retries now show a re-run button on the pull request, and you can also re-run them with a new Origin API endpoint. The check reads as pending, and still blocks merge if it is required, until the app posts a new result. Apps receive a new repository.check_run.rerequested webhook, and check runs now include rerequested_by.
  • Fixes to pull request pages and sign-in. A brief problem with Cursor's session check no longer signs you out of pull request pages, and stack pages load faster and still open for large organizations.
  • Fixes to repository reads and merge errors. File and blame reads on repositories with long delta chains no longer fail. Merging branches with unrelated histories, and checking whether stacks with more than 200 pull requests can merge, now return clear errors instead of internal server errors.
  • Fixes to repository pages and links. Commit links in GitHub's URL format (/owner/repo/commit/<sha>) on Origin's git host now redirect to the repository page instead of failing. Pinned and Recently Viewed repository cards now update their source icon when a repository's mirror status changes, and the "Origin access is not enabled" error now points you to cursor.com/codebase to request access.

API

  • Manage Origin apps from the API. You can now create, list, view, and update your namespace's apps through the Origin API, and add or revoke the signing keys they authenticate with, so you can script app setup and rotate keys without the web UI. App responses also include the owning namespace, description, website, and default scopes. See the Origin API changelog for endpoint details.
  • Manage repository mirrors from the Origin API. You can now change a mirrored repository's mirror state, detach it from its upstream source to make it a native repository, and poll the transition job until it finishes, all from the Origin API. See the Origin API changelog for endpoints and required scopes.
  • Choose merge or squash when merging through the API. The Origin API's Merge Pull Request operation and the merge_pull_request MCP tool take an optional merge method, merge or squash. Leave it out to use the repository's default. If you ask for a method the repository doesn't allow, the error starts with the same sentence other forges use, so existing merge tooling recognizes it. See the Origin API changelog.
  • Service account keys can read repository settings. Repo-scoped team API keys for service accounts can now read repository settings, rulesets, and labels through the Origin API, so automation can audit rulesets and compare settings without a user token. Changing settings still requires a user. See the Origin API changelog for details.
  • More detail in Origin API webhooks. Installation objects in webhook payloads now include when the installation was last updated, comment-created events carry the new thread's created and updated times, and every event payload is now documented with a sample. See the Origin API changelog for field details.
  • Fixes to Origin API comments and webhooks. Creating an inline comment or review through the Origin API with a line range past the end of the file now fails with a clear error naming the file and its line count, instead of being accepted. Webhook events are no longer dropped when a brief database outage interrupts delivery; they are retried. See the Origin API changelog for details.
  • Manage owner and repository access from the Origin API. You can now list, grant, change, and remove the access that users, groups, and team groups have on an owner or a repository through the Origin API. Each grant shows its permission level. See the Origin API changelog for endpoints and required scopes.
  • Update repository settings from the Origin API. You can now change a repository's default branch, allowed merge methods, automatic branch deletion on merge, and visibility through the Origin API. Repository responses also include these settings, so you can read and update them programmatically. See the Origin API changelog for endpoint details.

Agents

  • The agent steps back when you take control of its desktop. When you take control of a Cloud Agent's desktop, the agent stops using the computer instead of competing with you, and the desktop is released as soon as a computer-use subagent finishes.
  • Faster starts when you name the default branch. Cloud Agents started with the repository's default branch named explicitly, as the desktop app does, now reuse the prebuilt environment, the same as starts that name no branch, so they reach the first response sooner.

Projects

  • Project coordinators and workers show the right state. A Project coordinator is now marked unread only after a turn in which it actually sent you a message, and the messages it sends you now appear in the agent's conversation history. Workers and subagents that finished stay marked finished when a new message or a resubmitted message interrupts the turn that reports their completion.

Pull requests

  • Pull requests from mirrored repositories open in the right place. When a repository is mirrored between GitHub and Origin, Cloud Agents now open the pull request on whichever side is the source of truth and tell you which one they used. If the agent can't create the pull request, it reports the reason instead of working around it with another tool. Pull request links for mirrored repositories now point to the GitHub, GitHub Enterprise, or Origin host where the pull request actually lives, and agents no longer target a commit SHA as the base branch.
  • Merge approvals show the exact commit an agent will land. When a Cloud Agent asks for approval to merge a pull request on an Origin repository, the request now shows the commit subject and message the agent chose, in both the mobile approval sheet and Slack. You approve the commit that actually lands in history, not just the pull request title.

Subscriptions

  • CI subscriptions on your default branch no longer wake an agent on every merge. When a cloud agent subscribes to CI on a repository's default branch, such as main, it now wakes once for the next finished CI result, and then the subscription closes. To have an agent keep watching a pull request's CI, have it subscribe to the PR's branch.

Environments

  • Fixes to environment setup. The environments list loads faster for teams with many environments. The numbered steps in the environment setup intro dialog now line up evenly.

Self-Hosted Machines

  • Fixes for Cloud Agents on self-hosted workers. Trying to attach to the workspace of an agent on a self-hosted worker now shows "Workspace runs on a self-hosted worker" right away, instead of waiting on a workspace that never becomes ready.
  • Fixes to self-hosted worker claims. Self-hosted workers that reconnect no longer stay stuck on a stale claim, so they can be picked up for new agents again. Claim rejections no longer show up as "No self-hosted workers available", and claimed workers that never received a message are now released after the idle timeout.

Slack

  • Set a team default Self-Hosted pool for Slack. Team admins can run @Cursor pool set <name> so @Cursor mentions run on that Self-Hosted pool without adding pool= each time. If the pool serves any repository, mentions that name no repository can still launch an agent there. @Cursor pool and @Cursor settings show the current default, and @Cursor pool unset clears it.
  • Fixed Slack follow-ups that reached an agent without the thread. When you mention Cursor in a Slack thread before its Cloud Agent has posted a reply, the agent now receives the whole thread, including forwarded messages and files, instead of only your latest message. Follow-ups sent from the follow-up dialog get the same full context.
  • Mentioning @Cursor in another team's agent thread now starts a new agent. When you mention @Cursor in a Slack thread where the agent belongs to a different Cursor team, Cursor now starts a new Cloud Agent for you. Before, it refused and suggested Team Followups, which can't share agents across teams. If a follow-up is still denied, the message now says the agent belongs to another team.
  • Fixes to launching agents from Slack. When a Slack launch fails, the reply now names the repository and the real cause, such as no access, repo not found, a disconnected source control provider, or GitHub being unreachable, and links to GitHub Status and a Try Again button for that last case. Mentioning Cursor several times in one thread no longer starts duplicate agents when an earlier launch is slow, and your saved default repo and branch now still apply after your Slack workspace is migrated.

Jira

  • More reliable Cloud Agents from Jira. Cloud Agents started from Jira keep working with the latest Jira agent protocol, and show as working instead of paused while they run. Each new request from Jira now starts its own conversation, follow-up messages on an active task no longer fail, and stopping an agent while it starts now pauses it once it launches.

Fixes

  • Fixes to Cloud Agent reliability. Stopping an agent no longer shows a spurious provider error, follow-ups no longer fail to restore the branch when untracked files are in the way, and unnamed terminals in environment.json now start. Plugin slash commands in follow-ups now wait briefly for a plugin that isn't ready yet instead of skipping it.

Web app

  • Fixes to the agents web app. Clicking an agent in the agents sidebar on cursor.com no longer bounces back to the previous agent, and back and forward restore the tab you had open. A new agent started from the web now shows its logs and conversation once it's created, instead of missing them.

Integrations

GitHub

  • Accurate GitHub connect errors. If connecting GitHub fails, the callback page now says whether you were rate limited, hit an authorization problem, or are missing an installation, and shows an error code and request ID instead of a generic rate-limit message.
  • Clear error past the repository limit. When a request needs access to more repositories than GitHub allows in one installation token, Cursor now explains the limit instead of asking you to reinstall the GitHub App.
  • Team marketplaces refresh from private repositories. Plugin marketplace refresh can now authenticate through a GitHub App installation owned by another team in the same organization, so marketplaces on child teams pick up updates from private repositories.

GitLab

  • Connect self-hosted GitLab to the right team. Connecting a self-hosted GitLab instance from the Integrations page now links the selected team, and its repositories populate there, instead of your default team.
  • Reliable host removal. Deleting a GitLab host now removes all of its repositories, even when webhook cleanup fails for some of them.

Bitbucket

  • Complete Bitbucket Cloud pull request lists. Listing closed or all Bitbucket Cloud pull requests now includes merged, declined, and superseded ones.
  • Faster Data Center registration errors. Registering a Bitbucket Data Center instance now fails within seconds when the host is unreachable instead of hanging through retries.
  • Bitbucket Cloud pull request diffs. Pull request diffs from Bitbucket Cloud now load when Bitbucket redirects the request.

Dashboard and admin

Admin

  • Admin Download team data exports from the dashboard. Enterprise admins now have a Data exports tab that lists the exports prepared for their team, with a download link for each archive. The tab stays available after a subscription ends, so you can still retrieve your team's data. Available on Enterprise plans.
  • Admin Confirm before turning off Security Agents or Approval Agents. Turning off Security Agents or Approval Agents in team settings now opens a confirmation dialog, where you can optionally share why you're turning it off.
  • Admin Build the MCP allowlist from your configured servers. In team MCP Configuration, the MCP Allowlist's Add Rule menu lets you enter a custom rule or pick one of your team's configured MCP servers, each marked Allowed by default, Already allowed, or Blocked by allowlist. For a URL MCP server, Choose Tools lists the server's tools, after you sign in if needed, so you can pick the tools the policy allows instead of typing their names.
  • Admin Confirm decisions on new Cursor models. On the Models page, choosing Enable now or Keep disabled (previously Block) for a new Cursor model in its 7-day review window now asks you to confirm. Pending rows show the date the model will be enabled automatically and no longer carry a New model label.

Analytics

  • Admin Redesigned team Analytics charts. The Active Users chart on the Analytics page now stacks CLI, Cloud Agents, and Bugbot activity with an All line. Every chart shares one color palette with a clickable legend below it. Hover a chart to download its data or copy the matching API curl command.

Members

  • Admin Team Admin and Org Admin are now labeled separately. Member lists, role pickers, CSV exports, and the organization member detail page now show team admins as Team Admin and organization admins as Org Admin.
  • Admin Members > Set Cap now warns when a higher limit still applies. If you set a member's spend cap below a higher group or team limit, the dashboard tells you to lower that limit instead. Before, the save went through with no message and the member's limit stayed the same.

SSO and SCIM

  • Admin SCIM keeps team-owned groups and organization access in sync with your identity provider. When a directory group maps to a group owned by a team, SCIM now adds its members to that team, including existing members who were missing from it.

Fixes

  • Fixed several issues in the Automations editor. User MCP servers are now dimmed in the tool picker, with a tooltip, when an automation runs as a service account. The model picker no longer shows raw model IDs while models load, and the remove button on MCP server rows is always visible.
  • Admin Fixed SCIM group mappings that changed the wrong memberships. Mappings no longer add users to extra teams or refill the root team when a directory group already maps elsewhere, and no longer remove members still covered by another mapped group. Full syncs now defer removals when directory data is incomplete.
  • Fixed issues with cloud agent environments. Members who can't create team environments can no longer save a new environment with Team scope. When an environment build fails because the repository's GitHub App belongs to another team, the builds page now shows a warning saying so.

Automations

  • Auto tier now saves in automations. Choosing Cost, Balance, or Intelligence for Auto in an automation's model picker now saves correctly, and cloud runs launch with the tier you picked. Before, every tier reopened and ran as Balance.
  • Model pickers no longer flash raw model IDs. The model pickers in the automation prompt editor wait for the model list to load instead of briefly showing internal model IDs.
  • Fixes to automation saves. Saving an automation on a self-hosted GitLab repository whose host uses a host-level service account no longer fails while setting up webhooks.

Security and approval agents

  • GitLab and Bitbucket reviews. Security Reviewer now reviews GitLab merge requests and Bitbucket pull requests, posting a summary comment and a status check on the change.
  • More runs in run history. The Security Agents run history now includes Security Reviewer runs that ran on Cursor-managed hosting. Click one to open its session view.
  • Confirm before turning off. Turning off Security Agents or PR Routing & Approval for your team now asks you to confirm, and you can optionally say why.

Models and Cursor Router

  • New model: Muse Spark 1.3. Muse Spark 1.3 from Meta is now available in Cursor.
  • Cursor Grok 4.6 sees images directly. Cursor Grok 4.6 now receives images you attach and images returned by tools as images, instead of working from text descriptions of them.
  • Large generated files no longer block reviews. Files marked linguist-generated in .gitattributes, such as lockfiles, are left out of Bugbot's size check even when they are larger than 2 MB. Bugbot reviews the rest of the pull request instead of skipping it as too large.
  • Bugbot dashboard dialogs close with Escape. The disable, bulk repository action, and cancellation dialogs in the Bugbot dashboard now close when you press Escape.
  • Azure DevOps fixes. When a pull request summary would exceed Azure DevOps's 4,000-character description limit, Bugbot posts the full summary as a comment and says so in a shorter description. Organizations whose tenant check names no tenant can now enable Bugbot instead of failing tenant discovery. When an Azure DevOps pull request event includes its head commit, Bugbot queues the review without first reading the pull request from Azure DevOps, so a brief Azure DevOps or Entra outage no longer fails the event.

Reviews

  • Bugbot explains when a pull request is too large to review. When a pull request changes more than 100,000 lines or 3 million characters, Bugbot skips the review and posts a comment saying the pull request is too large to review, suggesting you split it into smaller pull requests.
  • Bugbot keeps Security Reviewer reviews. When Bugbot reruns on a pull request, it hides only its own earlier reviews. Reviews from Security Reviewer now stay visible.
  • Fewer failed reviews from brief network errors. A brief network error while Bugbot rechecks the pull request just before posting no longer fails the run and restarts the review.

Learned rules

  • Learned rules work on Origin repositories. Bugbot now learns rules from merged pull requests on Origin repositories. Previously, learning failed on Origin pull requests.
  • Learned rules work for teams that restrict models. If your admin model settings block Bugbot's default learning model, learning falls back to another allowed model. If every option is blocked, Bugbot skips the learning run instead of failing, and you aren't billed for it.

Dashboard

  • Bugbot dashboard fixes. Tables no longer make the whole page scroll sideways on narrow or mobile screens, and clicking a section header to copy its link only responds on the label itself.