Cursor version history

The Cursor version history lists the Cursor releases and feature names that could be confirmed against Cursor’s changelog, release notes, and documentation, and the npm and PyPI registries, as of 12 October 2026. The page also gives a routine for turning any new Cursor release note into a tested team rollout. For everything else, Cursor’s own changelog and release notes are the record.

This page is for the developer or tech lead who owns a team’s Cursor setup: the committed rules, skills, hooks, and MCP servers, and any script that drives Cursor’s agent. Your onboarding doc still says “Background Agents”, a teammate mentions a command you have never seen, and three people run three different builds. Before anyone changes a setting, you need to know what changed, whether it touches your config, and how you would notice if an update made the agent worse.

What you get from this Cursor changelog

• The current version of the Cursor desktop app, CLI, and SDKs, and where Cursor numbers each.
• A table of renamed features, so your team docs and prompts use the names Cursor uses today.
• Every changelog entry from 2026-08-19 to 2026-10-06 and every weekly desktop build from 3.19 to 3.24, each with its date and source.
• A six-step routine and three copy-paste prompts that take a release note to a team rollout, with a pass signal you read without reviewing every diff.

Which Cursor version is current?

As of 12 October 2026:

Item	Current value	How to check
Desktop app (Cursor IDE and Agents Window)	Cursor 3.24, released 2026-10-09 (Cursor’s release notes)	The version your app reports, and Cursor’s release notes. Record it with the date: “Cursor 3.24, checked on YYYY-MM-DD”
Cursor CLI	v2026.10.08, a date version	agent --version, and the CLI changelog
Cursor for iOS	1.12.0, released 2026-10-01	Cursor’s mobile release notes
TypeScript SDK	@cursor/sdk 1.0.37, published 2026-10-07	npm view @cursor/sdk version
Python SDK	cursor-sdk 1.0.37, uploaded 2026-10-07; needs Python 3.10 or later	pip index versions cursor-sdk, or the package page on PyPI
SDK runtime	Node.js 22.13 or later (engines field of @cursor/sdk)	node --version
SDK documentation	cursor.com/docs/sdk/typescript. The README of 1.0.37 links cursor.com/docs/api/sdk/typescript, which redirects there	The SDK docs
Models	The Cursor Models pool is Grok 4.7, Grok 4.6, Grok 4.5, and Composer 2.5. Cursor’s help calls Grok 4.7 “Cursor’s flagship model”, and Cursor added it in the week of 2026-09-21. Cursor names no single default model, and its list has no GPT-6 model (both checked 2026-10-12)	Prices are on the models hub; the live list is in Cursor’s model picker

The SDKs have a public version history on npm and PyPI, which makes them the Cursor artifacts you pin in CI. For how Cursor’s update channels work and how to roll back a desktop build, see Cursor upgrade guides.

How does Cursor number its releases?

Cursor numbers a release differently for each product:

• Desktop app. Minor versions on Cursor’s release-notes page. Its Cursor IDE page says each entry “covers one minor version, including the patch releases that follow it”. The per-product pages list 3.19 (2026-09-02), 3.20 (2026-09-10), 3.21 (2026-09-17), 3.22 (2026-09-24), 3.23 (2026-10-01), and 3.24 (2026-10-09), about one a week. They do not document 3.12 to 3.18 (checked 2026-10-12).
• CLI. Date versions: v2026.08.26, v2026.09.15, v2026.09.22, v2026.09.28, and v2026.10.08 (the latest on 2026-10-12).
• SDKs. 1.0.x versions on npm and PyPI.
• Marketing changelog. Entries at cursor.com/changelog carried a version label through 3.11 (Side Chats and Conversation Search, 2026-07-10). No later entry carries one (checked 2026-10-12).

Cursor cites a patch number when a feature needs a minimum build: Enterprise teams need Cursor 3.21.9 or later for Projects. Record your full build, not only the minor version. Model launches such as Grok 4.7 and Claude Opus 5.5 appear in the web release notes and on Cursor’s blog, not in the changelog (pages 1 to 3 checked on 2026-10-12).

On 2026-09-26 cursor.com was unreachable from the writing environment, so the previous edition of this page named no desktop build and hedged most Cursor facts. On 2026-10-12 cursor.com answered, and the values on this page were verified that day.

Which Cursor feature names changed?

Old names survive in team wikis, prompts, and review checklists long after the product moves on. A prompt that asks for a “Background Agent” still works on a person, but it confuses a new hire searching Cursor’s docs.

If your docs say	Cursor’s current name	What to do in your repository
Background Agents	Cloud Agents. Cursor’s help: “In Cursor, we call these Cloud Agents” (checked 2026-10-12)	Rename in docs and prompts. See Cloud Agents and Automations
BugBot, BUGBOT	Bugbot, the spelling on cursor.com/docs/bugbot (checked 2026-10-12)	Rename in review checklists. See Bugbot
YOLO mode, Auto-Run	Run modes: Auto-review (the recommended default since 3.6, 2026-05-29), Allowlist, and Run Everything. They live in Settings › Agents › Approvals & Execution	Name the mode your team means. In Run Everything “Every tool call runs automatically”; its CLI command is /run-everything, and /auto-run is an alias. Cloud Agents do not use Run modes, and Cursor says Auto-review “is not a security boundary”
Ask Every Time, Run in Sandbox	Allowlist. Cursor’s Run modes page records that 3.5 deprecated Ask Every Time and folded Run in Sandbox into Allowlist	For Ask Every Time’s behavior, use Allowlist with an empty allowlist. For Run in Sandbox, use Allowlist with sandboxing enabled
.cursorrules as the way to configure Cursor	Rules: project rules in .cursor/rules/*.mdc, or AGENTS.md in the project root and subdirectories. Cursor’s help: “The .cursorrules file in your project root is legacy and will be deprecated.”	Move to project rules or AGENTS.md. A plain .md file in .cursor/rules is ignored, and Cursor reads CLAUDE.md like AGENTS.md and always applies it. See project rules
A different command name for the Cursor CLI	agent: agent starts an interactive session, agent --version prints the version, and agent update updates it (Cursor’s installation page, checked 2026-10-12)	Check scripts and CI against that page. Names that other tools use for their Cursor integration are those tools’ own spellings, so leave them
Max Mode as a setting every plan has	Cursor’s help: “Max Mode is available only on legacy request-based plans” (checked 2026-10-12)	Remove it from setup guides for current plans
SDK docs at cursor.com/docs/api/sdk/typescript	cursor.com/docs/sdk/typescript. The /docs/api/sdk/typescript URL in the 1.0.37 README redirects there	Link the current path in internal docs. See the Cursor SDK

Correction: the edition of this page dated 2026-09-26 told you to move SDK links from cursor.com/docs/sdk/typescript to cursor.com/docs/api/sdk/typescript. That direction was wrong. On 2026-10-12 the /docs/api/sdk/typescript URL redirected to /docs/sdk/typescript.

What changed in Cursor, and which of it is verified?

Each row below was read on a Cursor page or a package registry on the date in the last column. The site’s earlier, longer history of Cursor 0.x to 3.x could not be re-checked on 2026-09-26, so this page does not reproduce it; the full record is Cursor’s changelog. Rows dated 2026-08-28 or 2026-09-26 in the last column are records of those checks.

Date	Change	Source	Checked
2026-10-07	@cursor/sdk 1.0.37 on npm and cursor-sdk 1.0.37 on PyPI. Cursor’s SDK release notes list releases up to 1.0.35, so 1.0.36 and 1.0.37 appear only on the registries	npm, PyPI, Cursor SDK release notes	2026-10-12
2026-10-06	Changelog: Remote control for local agents. “You can now see and reply to the local agents running on your computer from the Cursor iOS app.” On by default, except in Enterprise organizations, where admins enable it at Org settings > Security and identity > Remote control. A Keep this computer awake setting is included	Cursor changelog	2026-10-12
2026-09-23	Changelog: Rollouts and Security Review. Two bots “for the last mile of shipping code”, both on Teams and Enterprise. Rollouts attaches a monitor to every pull request, writes a monitoring plan as a PR comment, and reports each environment as verified healthy, regression detected, or inconclusive. On a regression it can open a revert PR or hand off to a cloud agent, but it “does not merge or roll back on its own today”. Security Review posts one comment per PR on exploitable bugs, among them injection, authn and authz bypass, committed secrets, SSRF, unsafe deserialization, and vulnerable dependency changes, and skips draft PRs	Cursor changelog	2026-10-12
2026-09-22	@cursor/sdk 1.0.32 on npm and cursor-sdk 1.0.32 on PyPI	npm, PyPI	2026-09-26
2026-09-21	Grok 4.7 added to Cursor in the week of 2026-09-21 (“New model: Grok 4.7” in the web release notes). Cursor’s blog index dates its “Introducing Grok 4.7” post 2026-09-21. Cursor’s help calls Grok 4.7 “Cursor’s flagship model”. The previous edition cited a secondary source (MarkTechPost) for this row	Cursor web release notes, Cursor blog, Cursor help	2026-10-12
2026-09-10	Changelog: Cursor Projects. A project takes on “a feature, a migration, or a full app”, keeps context over months of work, and delegates tasks to subagents. Its coordinator agent “doesn’t write code itself” and runs “on its own computer in the cloud”. Subscriptions let a project watch a Slack channel, run on a schedule, or follow PRs. “Projects are available in beta and rolling out to all users starting today.” Cursor’s Projects page says Projects is available on all paid plans, including Enterprise, and that Enterprise teams need Cursor 3.21.9 or later	Cursor changelog, Projects docs	2026-10-12
2026-09-02	Changelog: Self-hosted machines. The entry describes how to “keep tool execution entirely in your own network” with My Machines (one laptop or VM) or Team pools (named worker queues that scale and can hibernate idle machines). The entry names AWS Lambda, Coder, Cloudflare, Daytona, Modal, Namespace, Vercel, and E2B, and says self-hosted workers support computer use on Linux and Mac	Cursor changelog	2026-10-12
2026-08-27	Changelog: Start from scratch, without a repo. “Cloud Agents no longer require a connected GitHub or other third-party SCM provider to get started.” Start from scratch in the repo picker creates an Origin repo; the entry also covers Create repo, a live preview port-forwarded to the browser, and publishing through a connected Vercel account	Cursor changelog	2026-10-12
2026-08-19	Changelog: Cloud Agents and Cursor Harness Improvements, besides /goal and subagents (the next two rows), three more items. Subscriptions: a Cursor Agent subscribes to an event source, and “Subscriptions are available for cloud agents only, for now”. Custom modes: any skill can stay pinned in the chat as a mode, through Alt+Enter or Use as Mode. Steering: “Follow-ups wait for the next tool call instead of cutting the agent off mid-action.”	Cursor changelog	2026-10-12
2026-08-19	/goal: “give the agent a long-lived objective to work towards until it’s fully complete”. Marked “Rolling out” in the CLI slash-command reference, and still marked so in the agent overview and the CLI reference on 2026-10-12	Cursor changelog, agent overview, CLI reference	2026-10-12
2026-08-19	Subagents can run on their own virtual machines, each with an isolated copy of the project and clean context	Cursor changelog	2026-08-28
2026-05-20	Cursor 3.5 (changelog entry shared-canvases): /loop runs a prompt repeatedly on a local schedule until an outcome is reached or you stop it. It is a bundled skill	Cursor changelog	2026-08-28
2026-05-13	First uploads of the Python cursor-sdk to PyPI: a 0.0.0 placeholder, then 0.1.0 the same day	PyPI	2026-09-26
2026-04-26	First release of @cursor/sdk on npm (1.0.7)	npm	2026-09-26

For how /goal and /loop compare with the same commands in Claude Code and Codex, see goal commands across tools and loop commands across tools. Cursor lists /loop as a built-in skill that “Runs a prompt or skill repeatedly at a specified interval” (skills page, checked 2026-10-12).

Which builds did Cursor’s release notes list since late August 2026?

The release notes at cursor.com/docs/release-notes number the Cursor IDE and Agents Window builds, and list the CLI, web, Cloud Agents, Bugbot, mobile, and SDK releases on their own pages. Every row below was read there on 2026-10-12. Quotes are Cursor’s bullet titles.

Date	Release	Highlights
2026-10-09	Cursor IDE 3.24 and Agents Window 3.24	“Goal, queue, debug, canvas, and browser panels now open from pills above the prompt”. “Auto-review stops risky deletes for your approval”. A managed policy, AgentRemoteWorkspaceAuthorities: “Admins can restrict Agent to approved remote workspaces”. “Claude Opus 5.5 and Sonnet 5.5 on AWS Bedrock” (bring your own key). A Chat Width slider
2026-10-08	Cursor CLI v2026.10.08	Auto-review is stricter on risky deletes. Claude Opus 5.5 and Sonnet 5.5 on Bedrock through agent bedrock. /skills flags skills that break the Agent Skills limits, for example a name over 64 characters
2026-10-01	Cursor IDE 3.23, Agents Window 3.23, and Cursor for iOS 1.12.0	Approval prompts explain why they appear. “Privacy mode now blocks automatic crash debugging uploads”. A Workspace read-access mode in Run modes. Side chats on cloud agents. # references pull requests
Week of 2026-09-28	Cloud Agents, Bugbot, web, and CLI v2026.09.28	“Environment builds are always on”: they can no longer be turned off. “Start an agent from a closed pull request”. Bugbot: “Configure Bugbot per repository”. Web: “New model: Claude Sonnet 5.5”, and “New models: GLM 5.3 and GLM 5.3 Flash”, off by default
2026-09-24	Cursor IDE 3.22 and Agents Window 3.22	“Auto-review no longer blocks drafts and previews”. Hooks get parent_tool_call_id and child_conversation_id to tie subagent calls back to their parent. “Grok 4.7 now names itself correctly”. Pin up to 75 agents. A usage ring in the chat status bar
2026-09-22	Cursor CLI v2026.09.22 and Cursor for iOS 1.10.0	CLI: “Auto-review weighs block instructions against what a call does”, and fixed model instructions for Grok 4.7 and Opus 5.5. iOS: Projects and iPad chat
Week of 2026-09-21	Web and Cloud Agents	“New model: Grok 4.7”. “New model: Claude Opus 5.5”. “Auto follows team model settings”. “Long-running mode removed from the web”: Cloud Agents started from the web always run in standard mode
2026-09-17	Cursor IDE 3.21, Agents Window 3.21, and Cursor for iOS 1.9.0	Team rules with file patterns apply only to matching files. Separate browser sessions per workspace. Attach PDFs, code, and archives to cloud agents. Origin pull requests open in the Agents Window
2026-09-15	Cursor CLI v2026.09.15	“Asking the agent to stop a goal closes it”. “Loop reminders on every model”
2026-09-10	Cursor IDE 3.20 and Agents Window 3.20	“Agents can use the browser by default”. “Claude Fable 5.1 makes smaller, more focused edits”. Goals and background tasks keep the chat’s Auto-review setting. MCP, plugin, rules, and hooks settings open in Customize
Week of 2026-09-07	Web	“New model: Muse Spark 1.3”, described as the first Meta model available in Cursor. “Cursor Grok 4.6 sees images directly”. A 7-day review window for new Cursor models, with Enable now or Keep disabled
2026-09-02	Cursor IDE 3.19 and Agents Window 3.19	Stopping a subagent stops its nested subagents. /summarize on cloud agents. “See and manage what your cloud agent is listening for”
Week of 2026-08-31	Web	“New model: Claude Fable 5.1”. “New model: Gemini 3.8 Flash”. Self-hosted Machines pools in the dashboard. A Runs page
2026-08-26	Cursor CLI v2026.08.26	Persistent sessions: agent persist, /detach, agent persist attach, agent persist list, agent persist stop, and agent persist --resume. Computer use on macOS workers (agent worker --computer-use start). Team model restrictions apply in the CLI

Which Cursor features did the documentation describe in August 2026?

This inventory is what Cursor’s documentation described on 2026-08-28. Use it to check whether your team’s docs mention a feature under the right name and link to the right page on this site.

Area	Feature	What it does	On this site
Plan and build	Plan Mode	Creates an implementation plan before writing code	Agent modes
Parallel work	Worktrees, Subagents	Isolated Git checkouts for the agent; specialised assistants the agent delegates to	The Agents Window
Cloud	Cloud Agents, Builds, Automations, Cloud Agents API	Agents in isolated cloud VMs; background environment preparation; agents started on a schedule or by Slack, source-control, webhook, Linear, Sentry, or PagerDuty events; programmatic launch	Cloud Agents and Automations
Customisation	Rules, Agent Skills, Hooks, Plugins, MCP	System-level instructions; portable skill packages; processes that observe, block, or modify the agent loop over JSON on stdio; bundles of all of these; connections to external tools	Custom rules and templates
Review and release	Bugbot, Security Agents, PR Routing & Approval, Origin	PR review for bugs and security issues; security scanning; reviewer assignment and approval of low-risk PRs; Cursor’s git forge	Rollouts and approval
Scripting	CLI print mode (-p, --print), TypeScript and Python SDKs	Non-interactive runs for scripts and CI; Cursor’s agent called from your own code	The Cursor SDK

This inventory stays as the record of 2026-08-28. Features added after it, among them Self-hosted machines, Projects, Rollouts, and Security Review, are in the tables above.

How do you turn a Cursor release note into a tested rollout?

Treat a Cursor update like a dependency bump: one owner, one written note, and a pass signal that comes from tests rather than from someone reading the agent’s diffs. The tech lead who owns the team’s Cursor config signs off in step 6.

1. Record what each person runs today. Ask the team to post the build their app reports, and write it down with the date. If you script Cursor, record the SDK version your project actually installs:

   npm ls @cursor/sdk
   pip show cursor-sdk

2. Read every entry on Cursor’s changelog and on its release-notes page since your last check, and sort each one into three buckets: changes a default, a permission, or billing; touches config you commit (rules, skills, hooks, MCP servers, plugins); new capability you can ignore for now. The first prompt below does the sorting. To see how far behind your pin is, check the latest published version, to compare against:

   npm view @cursor/sdk version
   pip index versions cursor-sdk

3. Check that your committed config still loads. Open a fresh agent chat on the updated build and run the second prompt below. Compare its answer with the files in your repository: a rule or MCP server that is missing from the answer is not loaded.

4. Run a canary set. Keep three to five real tasks from your backlog, each with acceptance tests written before the agent starts. Run them on the old and the new build, each in its own worktree, from the same commit. The third prompt below is the template. Your pass signal is the test suite, the type check, and the linter, not the diff.

5. Pin anything that runs unattended. In CI, install an exact SDK version and move the pin only in a pull request that also runs your headless job on a fixture:

   npm install --save-exact @cursor/sdk@TESTED_VERSION
   pip install cursor-sdk==TESTED_VERSION

   Replace TESTED_VERSION with the version your canary run passed on. The latest published version on 12 October 2026 was 1.0.37. Do not copy a version from a registry page: the registry moves ahead of what you tested.

6. Roll out in two stages. One volunteer uses the new build for a few working days; then the whole team moves. The owner approves the second stage when the canary results match the old build and the volunteer has no open config issue. If they do not match, follow the rollback steps in Cursor upgrade guides before you debug.

Copy-paste prompt: triage Cursor release notes for this repository

Below are Cursor changelog entries published since our last check. Paste them
verbatim after this prompt.

First, read this repository's Cursor configuration: everything under .cursor/,
any rules, skills, hooks, MCP server configs, and plugins, plus scripts or CI
jobs that call the Cursor CLI or @cursor/sdk.

Then build a table with one row per changelog item:
- item (quote the changelog sentence, do not paraphrase it)
- bucket: DEFAULT/PERMISSION/BILLING change, TOUCHES OUR CONFIG, or OPTIONAL
- the files in this repository it affects (path), or "none"
- what we must test before rolling it out, as a concrete check

Do not invent features, version numbers or settings that are not in the pasted
text. If an item is ambiguous, mark it UNCLEAR and say what we should read.
Do not edit any files.

The prompt asks for the changelog sentence behind every row, so the reviewer checks each claim against its source instead of trusting a summary.

Copy-paste prompt: confirm what the updated agent loaded

Before doing any work, list exactly what you have loaded in this session:
1. Every rule, with its file path and why it applies here.
2. Every skill you can use, with its name and source directory.
3. Every MCP server and the tools it exposes.
4. Every hook or subagent definition you know about from this project.
Then compare that list with the files that exist under .cursor/ and any other
Cursor config in this repository. Report files that exist but are NOT loaded,
and anything loaded that does NOT come from this repository.
Do not change any files.

The agent’s list is a self-report; for a rule that must load, add a line asking the agent to start replies with a unique marker and check for it.

Copy-paste prompt: run one canary task after a Cursor update

Canary task. Implement the change described in docs/canary/rate-limit-headers.md.
The acceptance tests are in tests/canary/rate-limit-headers.test.ts and are the
specification: do not edit, skip or weaken them.
Work only in src/middleware/. When you finish, run:
  npm test -- tests/canary
  npm run typecheck
  npm run lint
Stop when all three pass or after three failed attempts.
Report: each command, its exit code, the number of files changed, and any
assumption you made that the task file does not state.

Replace the task file, test path, and commands with one of your own canary tasks. Run the same prompt on both builds and compare the report: exit codes, attempts used, and files changed. A new build that needs more attempts or touches more files on the same task is a regression signal even when the tests pass.

What breaks after a Cursor update, and how do you recover?

• The agent ignores a rule it followed yesterday. Run the “confirm what the updated agent loaded” prompt. If the rule is missing, compare its file against the current Rules page and move the team back to the previous build while you fix it; the rollback steps are in Cursor upgrade guides.
• An MCP server or hook stops working after the update. Check its entry against Cursor’s MCP or Hooks page, restart the server, and re-run the loaded-config prompt. Keep the server out of the rollout until the canary set passes with it enabled.
• A CI job that uses the SDK fails after an unpinned install. Pin the last version that passed (npm install --save-exact @cursor/sdk@<LAST_GOOD_VERSION>), then read the SDK release on npm before you move the pin again. On an older runner, check that Node.js is 22.13 or later first.
• A model you relied on disappears from the picker, or its behaviour changes. Cursor’s model list marks older models, for example Claude Opus 5, Sonnet 5, and Fable 5, as “Hidden by default” (checked 2026-10-12), so a missing model may be hidden rather than removed. Do not switch on a vendor benchmark table. Pick a candidate from the models hub, run the canary set on it, and follow the new-model playbook.
• The canary set passes, but with more attempts or larger diffs. Hold the team rollout. Reproduce on the previous build from the same commit, and report the difference to Cursor with both build numbers.
• A teammate cites a Cursor feature you cannot find in the changelog. Look in the release notes too: model launches appear there, not in the changelog. Treat it as unconfirmed until you read it on a Cursor page. Many third-party summaries of Cursor releases circulate; this page leaves out anything it could not trace to one.
• Team docs still use old feature names. Search the docs and prompts for the left-hand column of the renamed-features table and replace each hit in one pull request.

Where to go next after a Cursor update

Cursor reference hub The other Cursor reference pages

Cursor upgrade guides Update channels, pre-flight checks, and rollback

Feature adoption strategies Decide which new Cursor features your team adopts, and when

The Cursor SDK Drive Cursor's agent from TypeScript or Python

Reading evidence instead of code Pass signals you can trust without reading every diff

Models hub Current models, prices, and defaults across tools

The sibling changelogs track the other two tools the same way: Claude Code version history and Codex version history.