Cursor App
6 days since its last release · 3.28 ·
Cursor is an AI coding tool from xAI. Its latest release is 3.28, released on 6 Nov 2026. It shipped 5 releases in the last 30 days, and we have tracked 254 since 14 Mar 2023.
- Latest
- 3.28
- 6 Nov 2026
- Last 30 days
- 5
- releases
- Usual gap
- ~5d
- between releases
- Tracked
- 254
- since 14 Mar 2023
- Homepage
- cursor.com/changelog
Releases per month
last 12 months- 3 Dec
- 3 Jan
- 5 Feb
- 6 Mar
- 8 Apr
- 10 May
- 8 Jun
- 5 Jul
- 5 Aug
- 3 Sep
- 4 Oct
- 2 Nov
Release history
254 releasesReleases 100
-
Talk to your agents in realtime voice calls
- New voice chat lets you describe changes out loud, ask what an agent is doing and hear progress updates as it works
- Works in the Agents Window, the IDE, on web and in Cursor for iOS, started from a waveform button next to Send
- On iOS, say "Hey Siri, call Cursor" to start a call and mute or hang up from the Lock Screen
- Runs on Grok Voice with several voices to choose from, and zero data retention when Privacy Mode is on
- Separate from dictation (formerly Voice Mode), which still turns speech into text and is toggled with Cmd+Shift+Space
1 patch · latest 3.28.7
- 3.28.7
-
Build and deploy live apps from a cloud agent, shareable by link
- Cloud agents can now build web apps, set up a Postgres database, add auth behind Cursor SSO and handle hosting to produce a live link
- Security checks run before publishing to catch exposed secrets and credentials, and nothing goes live until they pass
- Team admins can view all apps and secrets, add review rules, approve blocked publishes and unpublish any app
- Apps can be shared with specific people, groups or the whole team, and shared publicly on individual plans
- Available on all paid plans with cloud agents enabled; deploys from Cursor Origin and GitHub repos, with GitLab coming soon
2 patches · latest 3.29.15
- 3.29.15
- 3.29.9
-
Agent SDK launches for building, testing and deploying custom agents
- Agent SDK is generally available: define an agent as a folder of code, test it with evals and deploy it to Cursor or your own infrastructure
- agent-sdk deploy builds the agent, gives it a URL and runs it on Cursor, with secrets and integrations managed from the dashboard
- Agents wake on GitHub, GitLab and Bitbucket events, Slack mentions and DMs, webhooks or a cron schedule, and follow-ups resume the same session
- Evals check which tools ran and the shape of the reply, and github replay turns a real pull request into test payloads
- Templates include a security reviewer, owners-policy approver, PR autofixer and Linear or Jira triager
4 patches · latest 3.27.22
- 3.27.22
- 3.27.14
- 3.27.12
- 3.27.6
-
Cloud agents can now build shareable interactive canvases
- Cloud agents can build interactive canvases while they work, and you can share a finished canvas with your team by sending its link
- New Environments API in beta lets you create and manage cloud agent environments from code
- Cloud agent environments now support up to 300 repositories
- Large environments load 5x faster
2 patches · latest 3.26.16
- 3.26.16
- 3.26.13
-
Origin App Marketplace adds 20+ integrations, opens to public submissions
- Origin now integrates with over 20 apps for deployments, CI, code review, security, monitoring and project management
- Partners include Cloudflare, Vercel, Netlify, CircleCI, Buildkite, CodeRabbit, Greptile, Semgrep, Socket Security, Sentry and Linear
- The marketplace is now open to public submissions, so any team building developer tools can publish an app after review
- Apps are built against the Origin API and can receive webhook events and write back check runs, pull request comments and reviews
- The Origin Apps plugin can answer questions from the live API docs and plan a port of an existing app to Origin
-
See and reply to your local agents from the Cursor iOS app
- New Remote Control lets you watch and message agents running on your computer from the Cursor iOS app
- Agents keep running on your computer, which must stay on and online; cloud agents are not required
- Pair a computer by tapping it in the iOS app and approving the request in the Cursor desktop app
- New Keep this computer awake setting stops the computer sleeping while you're away, if it is plugged in with the lid open
- On by default except for Enterprise orgs, where admins can turn it on in Org settings
-
Two new bots: Rollouts watches deploys, Security Review checks PRs
- Rollouts writes a monitoring plan on each PR and reports change health per environment as healthy, regression or inconclusive
- On a regression, Rollouts notifies the author and can open a revert PR or hand the fix to a cloud agent
- Security Review posts one comment per PR on exploitable bugs like injection, auth bypasses, leaked secrets and SSRF
- Each security finding includes severity, attack path and a proposed fix, and teams can add their own rules
- Available on Teams and Enterprise plans from the automations tab, with 10 days of free Rollouts credits
-
Projects arrive in beta for long-running work led by a coordinator agent
- New Projects take on larger work like features, migrations or full apps, keeping context over months of work
- A coordinator agent plans the work and hands it to many subagents running in parallel, then brings finished work back for review
- Projects run on their own cloud computer and keep going when your laptop is closed, starting local agents when testing on your machine is needed
- Shared context files sync across machines so agents keep what they learn about the codebase and your preferences
- Subscriptions let the coordinator watch Slack channels, run on a schedule or follow your PRs without being prompted
-
Self-hosted machines keep agent tool execution inside your own network
- Self-hosted machines run agent tool calls on your own infrastructure, so code, build outputs and secrets stay internal
- My Machines connects a single laptop or VM to your account for personal workflows
- Team pools are named worker queues that scale with demand and can hibernate idle machines, then restore them when a follow-up arrives
- Cloud agents can run on sandboxes you already use, including AWS Lambda, Coder, Cloudflare, Daytona, Modal, Namespace, Vercel and E2B
- Self-hosted workers on Linux and Mac support computer use, and you can watch the agent's desktop or take control from Cursor
-
Cloud Agents can now start from scratch without a connected repo
- Cloud Agents no longer need a connected GitHub or other SCM provider: pick Start from scratch in the repo picker and prompt right away
- Save the agent's work to a new Cursor Origin repo with the Create repo button, choosing a name and private or internal visibility
- Live in-browser preview of the cloud agent's environment via port forwarding, with support for tools like design mode
- Publish builds to a live URL by connecting a Vercel account
-
Cloud agents can now wake on PRs, Slack threads and schedules
- Cloud agents can subscribe to PRs, Slack threads or scheduled tasks and wake up when something happens
- Cloud agents follow the PRs they create through to completion, fixing CI and addressing bot comments
- New /goal command gives the agent a long-lived objective to work on until it is fully complete
- Subagents can run on their own virtual machines, each with an isolated copy of the project
- Any skill can be pinned as a Custom Mode, and follow-up messages now steer the agent without interrupting it
-
Origin brings code hosting to Cursor, with repos, pull requests and GitHub sync
- Origin code hosting begins rolling out in early beta on all paid plans, with repos, pull requests, code browsing and GitHub sync
- New Codebase tab to create Origin-hosted repos and push or clone them using the CLI
- GitHub repos can be synced in real time, and pull request comments, reviews and merges sync both ways
- Agents work inside every repo: they can answer questions about code, make changes, update PRs or push branches
- App integrations with Vercel for preview deployments and Depot or Buildkite for CI, plus per-repo access and sync settings
-
Cloud Agents now start from prebuilt environments, booting 3x faster
- New Builds feature prepares ready-to-use copies of your dev environment in the background, at no extra cost with Cloud Agents
- Environments boot 10x faster internally, with 3x faster time to first token
- If a commit or dependency update breaks a build, agents keep using the last successful one and you get notified
- New Builds tab in the Cloud Agents dashboard shows build status, logs, commit SHAs and which build each agent run used
- New environments use Builds automatically; existing ones can turn them on with Enable Builds or test the switch with a setup agent
-
Google Workspace plugins connect agents to Gmail, Drive and Calendar
- New plugins give coding agents direct access to Gmail, Google Drive and Google Calendar
- Google Drive: search files and folders, open and download content, create and organize files
- Gmail: search and read mail, draft and send messages, apply labels and manage threads
- Google Calendar: read schedules, create and update events, find free time
- Install from the Cursor Marketplace or the Customize page in Cursor
-
Cursor arrives on iPad, with a new inbox and full PR review on mobile
- Cursor for iPad is now available on all paid plans, with a layout built for the bigger screen
- Pinned sidebar chats let you watch several agents at once, and split screen keeps a review open next to a chat
- On iPhone and iPad, PR review now covers comments, checks, approvals and reviewers, and you can prompt the agent to resolve comments
- New Inbox shows what's in progress, what needs your attention and which PRs are in review
- Adds Bitbucket and Azure DevOps support, multi-PR sessions, team switching and Apple Pencil markup on screenshots
-
New ₹649 monthly Start plan for developers in India with UPI payment
- Cursor Start is a new ₹649 per month plan for India, tax inclusive, billed in INR with UPI or card
- Includes generous access to Grok and Composer models, with enough usage to build with agents daily
- Comes with always-on cloud agents and Cursor for iOS with remote control of agents
- Supports plugins, MCP servers, hooks and skills
- Free users in India can upgrade from the dashboard, and new users can choose Start at signup
-
Auto mode now runs on Cursor Router, which picks a model for each request
- Auto mode now uses Cursor Router, which sorts each request by task type and complexity and sends it to a frontier or price-efficient model
- New Cost, Balance and Intelligence modes set how the router trades quality against token spend
- Admins can turn the router on per team or group, limit modes, set the default and allow or block models
- Works on desktop, web, iOS, CLI and the SDK, and is on by default for Teams plans
- The routed model is hidden by default and can be shown, and admins can softly or strictly require Auto
-
Cursor in Slack now shares a plan first and works across multiple repos
- Cursor in Slack replies with a plan before starting and updates its status as it works, so you can redirect it early
- Can start in a named multi-repo environment, choosing the one that gives access to every repo the request needs
- A Switch repository button lets you move to another repo or environment mid-task, and Cursor picks up where it left off
- Can read from and post to other Slack channels and threads to pull in context and share updates
- Cleaner responses: in-message buttons replaced by compact footer links, and tables, PRs and artifacts render better
-
Side chats for tangents, plus search across agent conversations
- Side chats via /side, /btw or the plus button let you ask questions and explore tangents without interrupting the main agent
- Search agent transcripts from the command palette in the Agents Window, backed by a local index, and within a chat using Cmd+F
- Redesigned project and repo pickers with scoped search and in-picker GitHub, GitLab and Azure DevOps connection
- New Run on picker and a single Remote Machines menu, with working without a repo now an explicit No Repo choice
- New cloud agent hooks for prompts, responses, thinking, subagents, compaction and turn completion
-
Team marketplaces now support Team MCPs and organization groups
- Admins can configure Team MCP servers once and distribute them across cloud agents, the agents window, IDE and CLI
- Team MCP servers can be published to a team marketplace from Dashboard -> Plugins & MCPs so members install approved integrations without setup
- Team marketplace access can now be restricted to specific organization groups, alongside team-level SCIM directory groups
- Marketplaces already using SCIM directory groups keep their existing configuration
-
Cursor arrives on iOS in public beta to launch and steer agents from your phone
- iOS app in public beta on all paid plans: pick a repo and launch cloud agents with any frontier model, voice input and slash commands
- Remote Control lets you keep directing an agent running on your computer from your phone, with an option to keep the machine awake
- Live Activities on the lock screen and push notifications when an agent finishes, needs input or is ready for review
- Review demos, screenshots, logs and diffs on mobile, leave follow-ups or merge the PR directly from the app
- Teams and Enterprise admins must enable Remote Control from the Cursor Dashboard
-
Automations gain an /automate skill, new triggers and computer use
- New /automate skill creates an automation from your local agent session based on a plain-language description of the task
- Cloud agents started by automations can now use their own computers to produce demos or artifacts, enabled by default
- Five new GitHub triggers, including PR review comments, submitted reviews and completed workflow runs
- React to a Slack message with a designated emoji to kick off an automation
- Automations can be saved incomplete, open PRs by default and delete outdated memory files
-
Cloud subagents and guided cloud environment setup in Agents Window
- New /in-cloud command spins up a cloud subagent on its own VM and branch, keeping your local workspace clean
- Cursor can set up your dev environment in the cloud in under 10 minutes, saved as a reusable snapshot
- Commit the environment to .cursor/environment.json so cloud agents for your whole team start faster
- Use /babysit or the quick-action pill to have a cloud subagent prepare a PR for merge
- More reliable handoff of agent sessions between local and cloud, with unlimited parallel cloud agents
-
Bugbot gets over 3x faster and 22% cheaper, and finds more bugs
- Bugbot reviews now take about 90 seconds instead of about 5 minutes, cost about 22% less per run and find 10% more bugs
- Bugbot is now powered by Composer 2.5
- New /review command runs Bugbot and Security Review before you push, with /review-bugbot and /review-security as shortcuts
- Bugbot skips PRs whose diff was already checked with /review, on GitHub and GitLab
- Bugbot can now be set to review only what changed since its last review
-
Design Mode for canvases and an interactive context usage report
- Design Mode in canvases lets you select and annotate UI elements to guide edits instead of describing changes in text
- Context usage can be shown as an interactive canvas report breaking down tokens across system prompt, tools, rules and skills
- Debug with Agent button in the context report asks the agent to find ways to reduce context usage
- Shared canvases can be opened full-screen in the browser for presenting
- Agents can embed buttons in canvases that run a specific prompt when clicked
-
SDK adds custom tools, auto-review, custom stores and nested subagents
- Pass your own functions to the local agent as tools via local.customTools, available to every subagent in the run
- Route headless tool calls through auto-review, steered by natural-language allow and block instructions in permissions.json
- Persist agent state in a new JSONL store, or plug in your own store through the LocalAgentStore interface
- Subagents can now spawn their own subagents to any depth, with no setup required
- Scripts pinned to retired composer-2 move to Composer 2.5 automatically, plus fixes for wait(), HTTP/1.1 streaming and TypeScript types
-
Enterprise admins can now manage multiple teams under one organization
- Organizations are now generally available to all Enterprise customers as a top-level container for identity, administration and membership
- Run multiple teams, each with its own security, governance, spend and feature settings, and let users belong to several teams with different roles
- New groups give user cohorts separate model access, spend limits and agent permissions without creating a new team
- Organization-level identity provider management and usage analytics, with drill-downs into each team
- Admins can move users between teams through the dashboard, API or CSV, and new users inherit team settings automatically
-
New Auto-review run mode lets agents work longer with fewer approval prompts
- Auto-review is a new run mode that cuts approval prompts while keeping agent execution safer
- Covers Shell, MCP and Fetch tool calls: allowlisted calls run immediately and sandboxable calls run in the sandbox
- Other actions go to a classifier subagent that allows the call, tries a different approach or asks for your approval
- You can give the classifier custom instructions to steer its decisions
- Set your run mode in Settings > Cursor Settings > Agents > Approvals & Execution
-
Share agent-made canvases with teammates, plus a new /loop skill
- Canvases such as reports, dashboards and custom interfaces can now be shared with teammates as a link to a live snapshot
- Teammates open shared canvases in the browser with read-only access
- A new Shared Canvases page in the Cursor Dashboard lists and manages the canvases you have published
- Canvas sharing is available on Pro, Teams and Enterprise plans
- New /loop skill runs a prompt repeatedly on a local schedule, until an outcome is reached such as tests passing, or until you stop it
-
Cursor now works in Jira: assign tickets to start cloud agents
- Assign a Jira work item to Cursor, or mention @Cursor in a comment, to start a cloud agent
- The agent scopes the task from the ticket title, description, comments and your team's repository settings
- Ask it to fix bugs, add features, update tests or investigate an issue described in the ticket
- When the agent finishes, Jira shows completion updates and a link to the pull request
- Install it from Cursor integrations; requires Cursor admin access and Jira Commercial Cloud with Rovo enabled
-
Composer 2.5 model arrives with better long-task performance
- Composer 2.5 is now available, a substantial upgrade in intelligence and behavior over Composer 2
- Better at sustained work on long-running tasks and follows complex instructions more reliably
- Standard pricing is $0.50/M input and $2.50/M output tokens
- Fast mode is the default at $3.00/M input and $15.00/M output tokens
- Double usage included for the first week
-
Full-screen tabs and compact chat views in the Agents Window
- Files, changes, canvases, PRs, browsers and terminals can go full screen, with the chat replaced by a floating prompt bar (Cmd/Ctrl+Shift+M)
- Compact chat responses with a tool call density setting: Compact, Balanced or Detailed
- PR tabs show clearer review states, better reviewer and thread visibility, and more predictable actions
- Smoother long-chat scrolling, clearer status text for background tasks and faster search in large workspaces
- Fixes for background tasks, cloud agent setup, MCP/OAuth tokens, keyboard shortcuts and model routing
-
Delegate tasks to cloud agents from Microsoft Teams
- Mention @Cursor in any Teams channel to hand a task to a cloud agent or pull information from Cursor into Teams
- Automatically picks the right repository and model based on your prompt and recent agent activity
- Reads the entire thread for context, implements a solution and opens a PR for your team to review
- Install the integration from the Cursor dashboard
-
Bugbot gets adjustable effort levels for PR reviews
- Team admins and Individual plan users can set the effort level Bugbot uses for PR reviews
- Default effort works as before and is tuned for speed and efficiency
- High effort reasons for longer, costs more and may catch more bugs: 0.95 bugs per run on average versus 0.7 at default
- Custom mode lets you describe in natural language when Bugbot should use default or high effort
- Changing effort levels requires usage-based billing for Bugbot
-
New in-app PR review, parallel plan builds and PR splitting
- New PR review experience with Reviews, Commits and Changes tabs takes PRs from creation to merge in one place
- Build in Parallel runs independent parts of a plan at the same time with async subagents and keeps dependent steps in order
- New quick action splits changes into separate PRs, makes a backup snapshot and asks you to approve the split plan
- Pin your most-used skills as quick-action pills, and use /multitask in the editor to run async subagents
- Explore subagents can now use a chosen model, inherit the parent model or be turned off, plus fixes for MCP auth and cloud agents
-
Enterprise model blocklists, soft spend limits and richer usage analytics
- Admins get granular allow and blocklists by model and provider, including speed and context window configs, and can block new providers or versions by default
- Existing blocklists must be migrated to the new model controls system by June 1st via team model settings in the dashboard
- Soft spend limits no longer block users, and automatic alerts go out at 50%, 80% and 100% of soft or hard limits
- Usage analytics can be filtered by user or broken down by clients, Cloud Agents, automations, Bugbot and Security Review
-
Admins can create team marketplaces without a connected repo
- Admins can create a team marketplace without connecting a repository first
- Add, remove and configure first-party plugins directly in team marketplace settings
- Plugins bundle MCP servers, skills, subagents, rules and hooks to extend agents
- Three install modes: Default Off (opt in), Default On (opt out), Required (cannot uninstall)
-
Always-on Security Reviewer and Vulnerability Scanner agents in beta
- Security Reviewer checks every PR for vulnerabilities, auth regressions, data-handling risks and prompt injection, with inline comments on severity and fixes
- Vulnerability Scanner runs scheduled scans for known vulnerabilities, outdated dependencies and config issues, with findings sent to Slack
- Agents are customizable: adjust triggers, add instructions and plug in MCP servers for existing SAST, SCA and secrets scanners
- Beta on Teams and Enterprise plans; usage draws from your existing pool and admins enable it in the Cursor dashboard
-
Cursor SDK lets you build agents in TypeScript, plus Cloud Agents API rework
- New Cursor SDK (public beta) runs the same runtime, harness and models as Cursor's app, CLI and web agents, installed via npm install @cursor/sdk
- Agents can run locally on your machine or on Cursor's cloud against a dedicated VM, with any frontier model
- Billed at standard token-based consumption pricing; sample projects are available in a public repo and a native /sdk skill helps you start
- Cloud Agents API reworked around durable agents and per-prompt runs, with run-scoped follow-ups, status, streaming and cancellation
- Cloud Agents API adds SSE run streaming with Last-Event-ID reconnect, archive and delete controls, and standardized v1 response and error shapes
-
Async subagents, worktrees and multi-root workspaces in Agents Window
- New /multitask command runs async subagents in parallel instead of queueing requests, and splits larger tasks across a fleet of subagents
- Queued messages can be handed off to multitask instead of waiting for the current run to finish
- Improved worktrees in the Agents Window run isolated background tasks on different branches
- Any worktree branch can be moved into your local foreground with one click for testing
- Multi-root workspaces let one agent session work across several folders for cross-repo changes
-
Interactive canvases for dashboards and custom interfaces
- Cursor can now respond by creating interactive canvases, such as dashboards and custom interfaces
- Canvases are built from first-party components like tables, boxes, diagrams and charts
- Existing Cursor components such as diffs and to-do lists can also appear in canvases
- In the Agents Window, canvases are durable artifacts in the side panel next to the terminal, browser and source control
- Available in Cursor 3.1, in the Agents Window or the editor
-
New /debug mode, /btw side questions and /config panel in the CLI
- /debug starts Debug Mode: Cursor forms hypotheses, adds log statements and uses runtime info to find root causes before a targeted fix
- /btw lets you ask a side question without interrupting the agent's current run
- /config opens an interactive settings panel; /update-cli-config skill applies changes for you
- /statusline customizes the status bar with mode, branch, environment and task hints
- Image paste from clipboard, model picker from anywhere, auto-run suggestions, and fixes for Ctrl+Z backgrounding and lingering background shells
-
Tiled layout for parallel agents and better voice input
- Tiled layout splits the Agents Window into panes so you can run and compare several agents side by side; the setup persists across sessions
- Voice input transcribes your full clip in batch for better accuracy; hold Ctrl+M to speak, with waveform, timer and cancel/confirm controls
- Pick a branch before launching a cloud agent from the empty state, which avoids accidental runs on the wrong branch
- Jump from a diff straight to the exact line in the file, and use include/exclude filters in Search in Files
- Smoother performance: faster follow-ups in long chats, fewer dropped frames on large edits, and no more stutter when scrolling long conversations
-
Bugbot learns rules from PR feedback and gains MCP support
- Bugbot learned rules: turns reactions and replies on PRs into candidate rules, promotes useful ones and disables those that stop helping
- Bugbot can use MCP servers for extra context in reviews; Teams and Enterprise plans add tools in the Bugbot dashboard
- Resolution rate is now 78%, the highest to date
- New Fix All action applies multiple Bugbot fixes in one operation
- Autofix runs only on substantial findings, uses only relevant rules and has more reliable CI checks
-
New agent-centered interface for running many agents in parallel
- Agents Window runs many agents in parallel across repos: locally, in worktrees, in the cloud and over remote SSH
- Design Mode lets you annotate and target UI elements in the browser to give agents precise feedback
- Agent Tabs show multiple chats side-by-side or in a grid in the editor
- New /worktree and /best-of-n commands run tasks in isolated git worktrees, and cloud agents are removed from the Editor
- New Await tool lets agents wait on background commands, subagents or specific output; large-file diffs render faster
-
Self-hosted cloud agents keep code and tool execution in your network
- Cursor now supports self-hosted cloud agents that run entirely inside your own infrastructure
- Codebase, build outputs and secrets stay on your internal machines while the agent handles tool calls locally
- Same capabilities as Cursor-hosted cloud agents: isolated VMs, full dev environments, multi-model harnesses and plugins
- Enable it from the Cursor Dashboard
-
Composer 2 coding model now available, with Standard and Fast pricing
- Composer 2 is now available in Cursor, offering frontier-level coding performance on challenging coding tasks
- Standard pricing is $0.50 per million input tokens and $2.50 per million output tokens
- Fast is the default: $1.50 per million input tokens and $7.50 per million output tokens
-
30+ new plugins on the Cursor Marketplace from Atlassian, Datadog, GitLab and more
- More than 30 new plugins added to the Cursor Marketplace
- Partners include Atlassian, Datadog, GitLab, Glean, Hugging Face, monday.com and PlanetScale
- Cursor can read from, write to and take actions across more of your stack
- Most plugins contain MCPs that cloud agents can use when started manually or triggered by automations
-
Automations: always-on cloud agents triggered by schedules or events
- Build always-on agents that run on schedules or on triggers you define
- Triggers include Slack, Linear, GitHub, PagerDuty and webhooks
- Each run starts a cloud sandbox and follows your instructions with your configured MCPs and models
- A memory tool lets agents learn from past runs and improve with repetition
- Create automations at cursor.com/automations or start from a template
-
Cursor now works inside JetBrains IDEs via the Agent Client Protocol
- Cursor is available in IntelliJ IDEA, PyCharm, WebStorm and other JetBrains IDEs through the Agent Client Protocol (ACP)
- Use any frontier model from OpenAI, Anthropic, Google and Cursor for agent-driven development
- Install Cursor ACP from the ACP Registry inside your JetBrains IDE
- Sign in with your existing Cursor account
-
Interactive MCP Apps in chat and private plugin marketplaces for teams
- MCP Apps show interactive UIs inside agent chats, such as Amplitude charts, Figma diagrams and tldraw whiteboards
- Team marketplaces let Teams and Enterprise admins share private plugins with central governance and access controls
- Debug mode supports multiple parallel sessions, cleans up stale code each turn and adapts the number of instrumentation logs
- Browser tool gains a right-click context menu and zoom; multiple attached images now display as a gallery
- Many fixes to MCP configuration dialogs, review comment anchoring, diff display and long-running agent workflows
-
Bugbot can now automatically fix issues it finds in pull requests
- Bugbot Autofix runs cloud agents on their own machines to test changes and propose fixes directly on your PR
- Over 35% of Bugbot Autofix changes are merged into the base PR
- Bugbot comments on the PR with a preview of the fix, which you merge using the provided @cursor command
- Autofix can optionally push changes straight to your branch with no interaction required
- Enable it from your Bugbot dashboard
-
Cloud agents test their own work and show it with videos
- Cloud agents can now use the software they build to test changes and demo their work
- Each agent runs in its own isolated VM with a full development environment after onboarding onto your codebase
- Agents produce merge-ready PRs with artifacts such as videos, screenshots and logs for quick review
- Cloud agents are available on web, desktop, mobile, Slack and GitHub
-
CLI plan handoff to cloud and inline Mermaid diagrams
- After a plan is generated, the CLI shows a persistent menu to build in the cloud or locally; /plan returns to it and Shift+Enter builds in cloud
- Mermaid code blocks render inline as ASCII diagrams in the CLI; Ctrl+O switches between the diagram and its source
- Cursor Blame now tracks CLI AI code and keeps attribution when formatters or pre-commit hooks rewrite code
- Sandbox gets granular network access controls, and WebSearch and WebFetch share one domain allowlist
- Agent sessions are saved as JSONL transcripts, plus faster boot, background auto-updates and many UX and reliability fixes
-
Plugins marketplace, sandbox network controls and async subagents
- Plugins bundle skills, subagents, MCP servers, hooks and rules into one install, via the Cursor Marketplace or /add-plugin
- Sandbox gains granular network allowlists plus file and directory access controls; Enterprise admins can enforce egress policies
- Subagents can run asynchronously and spawn their own subagents, so the parent keeps working
- Plan mode can hand off to a Cloud Agent with Build in Cloud
- Agents can search past conversations; git clone, npm install and pip install work in the sandbox by default
-
Long-running agents arrive in research preview for bigger tasks
- Long-running agents plan first, then work autonomously over longer horizons on larger, more complex tasks
- Completed work that was previously too hard for regular agents in research preview and internal testing
- Produces larger, more complete PRs with fewer obvious follow-ups
- Available at cursor.com/agents for Ultra, Teams, and Enterprise plans
-
Subagents, Skills and image generation come to agents
- Subagents run in parallel with their own context; defaults cover codebase research and terminal commands, and custom subagents are supported
- Agent Skills defined in SKILL.md files now work in the editor and CLI, and can be invoked from the slash command menu
- Agent can generate images from text or a reference and saves them to the project's assets folder
- Cursor Blame (Enterprise) adds AI attribution to git blame, separating Tab, agent and human edits
- Agents can ask clarifying questions in any conversation while they keep working
-
Plan and Ask modes and cloud handoff come to the CLI
- Plan mode in the CLI via /plan or --mode=plan, with clarifying questions to refine your approach before coding
- Ask mode in the CLI via /ask or --mode=ask for exploring code without making changes
- Hand a local conversation off to a Cloud Agent by prepending & to a message, then resume on web or mobile
- One-click MCP authentication with automatic callback handling, plus an interactive /mcp list menu
- New hooks for session start/end, prompt and stop, plus WebFetch and WebSearch tools with approval options
-
New CLI model, rules, MCP and command controls; hooks 10x faster
- New agent command is now the primary CLI entrypoint; cursor-agent stays as an alias
- List and switch models with agent models, --list-models or the /models slash command
- Create and edit rules with /rules and commands with /commands from the CLI
- Enable or disable MCP servers on the fly with /mcp enable and /mcp disable
- Hooks run in parallel with merged responses and execution latency is down 10x
-
Stability fixes for agents and layouts, plus new switchable layouts
- Four default layouts (agent, editor, zen, browser) with Cmd+Option+Tab to switch, and Shift to go backwards
- Parallel agents and worktrees are more stable, with many fixes to accept/reject controls, model names and file deletions
- Extension processes are now separated from built-in Cursor extensions such as codebase indexing
- Integrated browser supports multiple tabs, and the update button appears less often
- MCP OAuth adds Static Client headers with scopes and state, and reopens auth tabs on startup
-
Enterprise insights, billing groups, service accounts and Linux sandboxing
- Conversation insights classify agent sessions by category, work type and complexity, without collecting PII
- Billing groups map usage and spend to your org structure, with budget alerts per group
- Service accounts with API keys let teams automate Cursor and cloud agents; Enterprise rollout starts week of 12/22
- Agent sandboxing now works on Linux as well as macOS, limiting network and filesystem access
- Share read-only agent transcripts in PRs or docs; others can fork them to start new conversations
-
Debug Mode, browser style editor, and multi-agent judging
- Debug Mode instruments your app with runtime logs to find the root cause of tricky bugs, across stacks, languages and models
- New browser sidebar with component tree and style editor to move elements, change colors and CSS, then apply changes via agent
- Multi-agent judging automatically evaluates parallel agent runs and recommends the best one with an explanation
- Plan Mode adds inline Mermaid diagrams and lets you send selected to-dos to new agents; plans are now saved as editable files
- Pinned chats in the agent sidebar, plus read-only terminal in Ask Mode and Rules that support prompts and scripts in folders
-
Smarter Plan Mode, in-editor AI code review and instant grep
- Plan Mode asks clarifying questions through an interactive UI, and you can search inside generated plans with Cmd+F
- AI code reviews find bugs in your changes and list the issues in the sidepanel, alongside Bugbot
- Instant Grep (beta) makes all agent grep commands and sidebar searches instant, rolling out over the next week
- Cloud Agent support added for GitLab, GitHub Enterprise, Sentry and Plain
- Custom modes removed (exportable as custom commands); project-level hooks now supported
-
Multi-agent editor and Composer, Cursor's first coding model
- Run up to eight agents in parallel on one prompt, each isolated in its own git worktree or remote machine
- Composer is Cursor's first agentic coding model, 4x faster than similarly intelligent models
- New editor with a sidebar for agents and plans, plus easier review of Agent changes across files
- Browser for Agent is now GA and embeddable in-editor; sandboxed terminals are GA on macOS and on by default
- Voice mode, Team Commands, plan mode in background, and faster LSP performance with fewer memory leaks
-
Browser controls, Plan Mode and hooks for longer agent runs
- Agent can take screenshots, improve UI and debug client issues through new browser controls
- Plan Mode writes detailed plans before complex tasks, letting agents run for significantly longer
- Hooks (beta) let custom scripts observe, control and extend the Agent loop, e.g. block commands or redact secrets
- Terminal commands now run in a sandbox with workspace access and no internet; failed commands can be retried outside it
- Team rules from the dashboard apply to all projects and Bugbot, which now also writes PR summaries
-
Custom slash commands, /summarize and a more reliable Agent terminal
- Create reusable custom slash commands stored in .cursor/commands and share them with your team
- New /summarize command frees up context window space on demand without starting a new chat
- Agent terminal is much more stable, fixing hangs and commands not exiting, and improving SSH
- Added MCP Resources support and interpolated variables such as environment variables in MCP server config
- Terminal UI is faster and polished, with OS notifications when shell commands need approval
-
Start Background Agents from Linear, plus OS notifications and MCP elicitation
- Start Background Agents directly from Linear issues to delegate tasks to Cursor
- Native OS notifications when an agent run finishes or needs input, enabled in Settings
- MCP elicitation support lets servers request structured input from users via JSON schemas
- Agent terminal opens on the left with a highlighted backdrop, and rejecting auto-focuses the input
- Jupyter: Tab sees the whole notebook, and edit notebook tool works with all models
-
Better Agent steering and tools, plus GitHub PR support for Background Agents
- Messages sent while the agent works now steer it at the next ideal time, usually after a tool call; Alt/Ctrl+Enter options queue or interrupt
- Improved Agent tools for large codebases: full-file reads with the 2MB cap removed, directory tree listing, better Grep and Codebase Search
- Tag @Cursor in GitHub pull requests and the Background Agent applies the fix and pushes a commit
- Assign a different model to each Agent tab, and manage all foreground and background Agents from the left sidebar
- Usage stats now show in Chat, with a new compact chat mode; Background Agents start twice as fast
-
Agents share your terminal, plus context usage view and faster edits
- Agents can now use your native terminal, running in the background until you click Focus to watch or take over
- Chat shows how much of the context window is used at the end of a conversation
- Faster edits: Search & Replace latency down 25% and Apply edits down almost 11%
- Auto-run security improved by replacing the denylist with allowlists; Manual mode removed from chat
- Checkpoints now work with Notebooks; team admins can configure an extension allowlist
-
Agent to-do lists, queued messages, PR search and faster Tab
- Agents plan long tasks with structured to-do lists, visible in chat and streamed to Slack
- Queue follow-up messages for Agent and reorder them while it works
- Memories is now generally available, with user approval for background-generated memories
- PRs are indexed and searchable; fetch a PR, issue, commit or branch into context
- New embedding model improves codebase search accuracy; Tab is about 100ms faster
-
Launch Background Agents from Slack by mentioning @Cursor
- Mention @Cursor in a Slack thread to start a Background Agent that reads the whole thread for context and opens PRs in GitHub
- Agents run remotely and post updates in Slack, with links to Cursor and GitHub when the work is done
- An admin must connect the integration first, from Dashboard → Integrations
- Use the help and settings commands in Slack to see commands and set default model, repo and branch
- MCP gains progress notifications, dynamic tool registration and roots support; settings are searchable with Cmd/Ctrl+F
-
Bugbot code review, Background Agent for all, one-click MCP install
- Bugbot reviews your PRs on GitHub and flags bugs, with a Fix in Cursor button that opens the editor with a pre-filled prompt
- Background Agent, the remote coding agent, is now open to all users via the cloud icon in chat or Cmd/Ctrl+E
- One-click MCP server setup plus OAuth support for authenticating servers
- Agent can create and edit multiple cells in Jupyter Notebooks, supported with Sonnet models to start
- Memories beta lets Cursor remember facts from conversations per project, enabled in Settings → Rules
-
Unified request pricing, Background Agent preview and Max Mode for all top models
- Pricing is now unified into request-based usage; Max Mode uses token-based pricing, and premium tool calls and long context mode are removed
- Max Mode is available for all top models, selectable from the model picker
- Background Agent preview runs agents in parallel in remote environments, enabled in Settings > Beta
- New Tab model suggests edits across multiple files, with syntax highlighting in suggestions
- Refreshed Inline Edit adds full file edits and send-to-agent; also @folders, multi-root workspaces, chat export and duplication
-
Generate rules from chats, review diffs, MCP images, global ignores
- Generate rules from a conversation with /Generate Cursor Rules; Auto Attached rules apply by path and Always rules persist in long chats
- Built-in diff view with a Review changes button at the end of each agent conversation
- MCP servers can now receive images such as screenshots, UI mocks and diagrams as context
- Agent terminal commands can be edited before running or skipped; Pop-out renamed Move to background
- Global ignore patterns in user settings, new models (Gemini 2.5, Grok 3, GPT-4.1, o3, o4-mini) and beta project structure in context
-
Chat tabs, custom modes and much faster indexing for shared codebases
- Chat tabs (Cmd+T) let you run several conversations in parallel, with an orange dot when a tab is waiting for you
- Custom modes (beta) let you build modes with your own tools and prompts; Edit is renamed Manual and Ask can search the codebase itself
- Indexing of similar codebases within a team is much faster, with the Cursor repo going from about 20 minutes to under a minute
- Usage-based models now show cost per chat and input token counts per message; chat warns you when nearing usage limits
- MCP is more reliable on Windows with better error messages; reject all diffs moves to Cmd+Shift+Backspace and auto-run is disabled for everyone
-
Stability and memory fixes, plus new shortcuts menu and early access opt-in
- MAX mode for Claude 3.7 Sonnet, with and without thinking; thinking now costs 2 requests instead of 1
- Lower memory usage and better stability, with faster applies and a warning when edits fail because of cursorignore
- All keyboard shortcuts now editable in Settings > Keyboard Shortcuts
- Opt in to early access via Settings > Beta > Update frequency; release track picker added
- Auto select model picks the best premium model for the task, even during outages
-
Agent becomes the default mode, plus UI refresh and web search
- Agent is now the default mode, replacing the separate Chat, Composer and Agent modes with one unified interface
- Agent can search the web automatically, without needing an explicit @Web command
- Agent can run MCP tools automatically with Yolo mode, use MCP resources as context, and read project servers from .cursor/mcp.json
- New default themes and a simplified @-context menu in the first phase of a UI refresh
- Hitting the 25 tool call limit lets you press continue, which counts as a new request
-
Project rules in .cursor/rules, DeepSeek models and better codebase search
- Write repository-level rules to the .cursor/rules directory and the Agent picks which rule to follow
- DeepSeek R1 and v3 models supported, self-hosted in the US, enabled in Settings > Models
- New Codebase Understanding model and a new Fusion Tab model, better at jumps and long context, rolling out shortly
- Agent can see your recent changes, and you can summarize previous Composers to continue in a new conversation
- Optional larger context window for long tagged files with premium models, using more fast requests
-
Agent gets Yolo Mode, background commands and linter-driven fixes
- Yolo Mode lets the agent auto-run terminal commands; it also sees exit codes, can run commands in the background, and commands are editable
- Agent reads linter errors to fix issues automatically, auto-saves changes and can edit multiple locations in parallel
- @docs, @git, @web and @folder now work in the agent, which also supports 4o and has a smarter apply model for reapplying edits
- Cursor Tab can make larger edits at a time
- Composer changes and checkpoints persist across reloads, with a better review UX and a cheaper, faster Bug Finding Model
-
New Composer sidebar UI with early agent mode and commit messages
- Composer now lives in the sidebar with inline diffs
- Early agent in Composer can pick its own context and use the terminal
- Git commit messages can now be generated
- File pill recommendations and @Recommended for semantic context search in chat and Composer
- Nicer image dropping, performance improvements, and a beta sneak peek of a bug finder
-
Composer history, model search and @Lint Errors return
- Composer history lets you reopen earlier sessions after a restart and edit and resubmit past messages
- Model search (Cmd-option-/) makes switching models easier in chat, composer and cmd-k
- @Lint Errors is back in Chat, with slight improvements to Debug with AI
- Cursor is now based on VS Code 1.93.1; Python auto import for Cursor Tab is more stable
- 0.42.1 to 0.42.5 fix CVE-2024-43601, add mitigations for CVE-2024-48919 and fix composer bugs
-
Python auto-import in Cursor Tab, Composer upgrades, Remote SSH to Mac
- Cursor Tab now auto-imports symbols in Python files and is more stable
- Composer can be added to the AI pane and now supports images
- Composer Notepads (formerly Projects) can include tagged files and be referenced in chat and composer
- Cursor can now be used on Macs over Remote SSH
- Apply and Composer are slightly faster; fixes security issue CVE-2024-45599
-
New chat UX, Composer on by default and a smarter Tab model
- New chat UX replaces the previous chat experience
- Composer is on by default for all Pro and Business users via cmd+I
- Composer Projects (beta) lets several composers share instructions
- New Cursor Tab model is smarter and more context-aware
- Auto imports (beta) for Cursor Tab in TypeScript files, enabled in Settings > Features > Cursor Tab
-
Faster Cursor Tab with chunked streaming, plus concurrent composers
- Cursor Tab (formerly Copilot++) now defaults to chunked streaming and includes several speedups
- Composer supports running multiple composers concurrently
- New composer control panel
- Fixed a bug where accepted files were deleted
- 0.39.1–0.39.6 fix Cursor Tab rendering bugs and hangs, plus an unresponsive file explorer
-
Faster chunked Copilot++ edits (beta) and Claude 3.5 Sonnet as default
- Copilot++ Chunked Streaming (beta) surfaces edits faster in smaller chunks; enable it under Features > Copilot++ in settings
- Claude 3.5 Sonnet is now the default model for users
- Composer gains a file picker, arrow key navigation and a model toggle, plus fixes for outstanding bugs
- Cursor is now based on VS Code 1.91.1
- 0.38.1 fixes a bug that migrated OpenAI API key users to Claude 3.5 Sonnet
-
Experimental multi-file editing with Composer (beta)
- New experimental Composer feature for editing across multiple files
- Enable it under settings gear, then Beta, then activate Composer
- Open Composer with Cmd+I
-
Instant code applies from chat and a new Docs re-indexing setting
- Click Apply on a code block suggested in chat to instantly see the change in the file (small enough files only)
- New docs management: re-index your docs from Cursor Settings > Features > Docs
- Bug fixes when using your own API key for Claude
- Follow-up fixes: cmd-shift-F on macOS x64, slow-to-appear official docs, and buggy cmd-K stickiness
-
Cursor Prediction on by default, plus remote tunnels and sturdier SSH
- Cursor Prediction is now on by default, with a new UI
- Remote tunnels are now supported
- Remote SSH is more robust, including support for multiple proxy jumps
- Chat messages show context pills so you can see what context will be or was used
- Copilot++ partial accepts are off by default with a configurable keybinding, and gpt-4o is the default model
-
VS Code 1.89 merge, new prediction UI and partial Copilot++ completions
- Merges VS Code 1.89 into Cursor
- New Cursor Prediction UI
- Accept partial completions with Copilot++, which also handles linter errors better
- Gemini 1.5 Flash available in long-context mode, and GPT-4o in Interpreter Mode
- Toggleable rerankers on codebase search
-
Fixes a connection error and adds Command-K region autoselect
- Fixes a connection error that consistently affected some users and improves performance on spotty internet
- Command-K now automatically selects the region you're working on; manual selection still works
- Follow-up builds fix settings toggles and Copilot++ diffbox performance, plus onboarding tweaks
-
Copilot++ syntax-highlighted previews, new GPT-4 Turbo model, .cursorrules
- Copilot++ suggestion previews now have syntax highlighting, making changes quicker to understand
- Repo-level AI rules: add a .cursorrules file in the repository root for context and style guidelines
- New gpt-4-turbo-2024-04-09 model, enabled in Settings > Models
- Cursor Help Pane (beta) answers questions about features and shortcuts, enabled in Settings > Beta
- Follow-up fixes: Copilot++ highlighting performance, slower Copilot++ over SSH, preview box, AI Notes off by default
-
New experimental Long Context Chat beta for chatting across many files
- Long Context Chat (Beta) lets you chat with lots of files; enable it under Settings > Beta
- Select Long Context Chat in a new chat and @ a folder or the entire codebase
- Fixes a bug where empty or partial responses showed in chat
- 0.31.1–0.31.3 restore AI Review (alpha), fix the Cursor Settings menu item and fix @web returning no response
-
Copilot++ is about 2x faster, and Claude models arrive for Pro and API users
- Copilot++ is about 2x faster thanks to a new model and faster inference, rolling out to everyone over a few days
- All newest Claude models are available to Pro and API key users, toggled on in Settings > Models
- Pro users get 10 free Claude requests a day, then continue at API-key prices
- Inviting teammates is easier, and team admins can mark themselves unpaid and see last usage
- All settings now sit behind the gear in the top right, replacing the More tab
-
Use Claude 3 Opus as a custom model for Pro and Business users
- Pro and Business users can add claude-3-opus as a custom model in Settings
- Includes 10 fast requests per day for free
- Slow requests are unlimited, but the delay increases exponentially
- A more permanent solution, including API key users, is expected very soon
-
AI Notes on by default, plus better in-editor chat and interpreter mode
- AI Notes is now enabled by default; hold shift on any symbol to use it
- Better in-editor chat
- Auto-execute interpreter mode added
- Nicer onboarding styling and feedback modal
- 0.29.1 fixes Copilot++ sometimes not showing suggestions, hint line covering ghost text, and AI Notes not working on Windows
-
Rebased on VS Code 1.86.2 with sticky scroll in tree views
- Cursor is now based on VS Code 1.86.2
- Sticky scroll now works in tree views
- cmdk prompt bars are now sticky
- 0.28.1 fixes a spacing issue with codebase chat
-
New AI linter for saves and a stronger Interpreter Mode backend
- New experimental AI linter, turned on in the More tab beside Chat, scans your file for small bugs every time you save
- Interpreter Mode backend greatly improved, with better tool use and understanding of your project
- Fixes to the Windows build, chat context UI and onboarding (0.27.1-0.27.4)
-
AI Previews beta for quick code notes, plus finer chat replies
- AI Previews beta: an experimental code reading feature, enabled in the More tab beside Chat; hold shift to see quick notes on the symbol you're on
- Fine-grained chat replies: hover over the part of a response you want to reply to
- Copilot++ shows ghost text more often, can be toggled from the status bar, and has a more visible suggestion box
- Smoother onboarding: fixed Windows settings import and added an option to import folder and window state
-
Cmd-I to Heal: AI fixes selected code, turning pseudocode into real code
- Hold cmd-I over a selection to heal the code with GPT-4
- Useful for writing pseudocode and having the AI convert it into correct code
-
Drag images into the Command-K prompt bar
- Command-K now supports vision: drag images straight into the prompt bar
- Search through your past chats
- Applying diffs from chat is a bit faster
- Copilot++ performance improvements (0.25.2)
- Fix for a Command-K bug (0.25.3)
Showing the newest 100 of 254.