The short answer
Cursor is an AI coding agent and code editor made by Anysphere, which became a wholly owned subsidiary of SpaceX on 14 August 2026. It reads your codebase, edits many files at once, runs terminal commands and can run agents in the cloud. Since 12 October 2026 it is sold as part of The X Subscription: paid plans include Grok 4.7 and Composer 2.5, while Claude, GPT and Gemini models cost extra and need Premium or above.
| Made by | Anysphere (Cursor), a SpaceX subsidiary since 14 August 2026, United States |
|---|---|
| Free plan | Yes, with limited Agent and Chat requests and no third-party models |
| Models | Grok 4.7, Grok 4.6, Grok 4.5 and Composer 2.5 included; Claude, GPT-5.6 and Gemini at API list price on Premium and above |
| Best at | Multi-file changes in real codebases, parallel agents, long-running cloud work |
| Shows sources | Not a research tool; it cites the files and lines it read in your code |
| Team plans | Teams and Enterprise, with SSO, team-wide Privacy Mode and admin controls |
| Your data | With Privacy Mode on, code is not used for training. With it off, Cursor may store and train on code data. On by default for Enterprise. |
What Cursor is in November 2026
Cursor started as a code editor with an AI chat panel. It is now closer to a workplace for coding agents. You can still open a file and ask for a change, but the centre of the product since Cursor 3.0 (2 April 2026) is the Agents Window, where several agents work in parallel: on your laptop, in separate git worktrees, in the cloud or on a remote machine over SSH. The desktop app runs on macOS, Windows and Linux, and there is a command line tool, a web version for starting cloud agents and an iOS app that, since 6 October 2026, lets you reply to agents running on your own computer.
Two business changes matter more than any feature. First, SpaceX completed its acquisition of Cursor on 14 August 2026, in an all-stock deal reported at about 60 billion dollars. Cursor said the deal gives it access to far more computing power to train its own models. Second, on 12 October 2026 Cursor became part of The X Subscription, a single subscription for Cursor, Grok and X. The new tiers are Free, Plus, Premium, Super and Ultra, plus Teams and Enterprise. Your usage now comes as one weekly pool shared between Cursor and the Grok apps, and it refreshes every week rather than every month.
The practical consequence is about models. The weekly pool covers Cursor's own models: Grok 4.7, which Cursor introduced on 21 September 2026 as its most capable model for long-running coding, Grok 4.6, Grok 4.5 and Composer 2.5. Third-party models such as Claude Opus 5.5, Claude Sonnet 5.5, GPT-5.6 Sol and Gemini 4 Argon are still in Cursor, but only on Premium and above, and they are paid from prepaid credit at each provider's API list price (Teams and Enterprise also pay a Cursor Token Rate on top). On Plus and Free you cannot use them at all. If you were on one of the older plans (Pro, Pro+, the old Ultra or a Teams Standard or Premium seat), Cursor says those legacy plans keep working as they are, including the included third-party usage, until you choose to switch.
What Cursor is genuinely good at
Changes that touch many files
The thing Cursor does better than a chat assistant is work inside your actual repository. Ask it to rename a concept across a codebase, add a field from the database through the API to the screen, or replace one library with another, and the agent searches the code, edits every file involved and runs your tests in the terminal. You review the result as a set of changes, not a pasted snippet you have to place yourself.
Comparing approaches side by side
The /best-of-n command runs the same task across several models, each in its own isolated worktree, and lets you compare the outcomes. This is useful when you are unsure which approach is right, or when one model keeps getting stuck. The /worktree command keeps an agent's changes isolated from your main working copy so an experiment cannot break what you are doing.
Long jobs that keep running without you
Cloud agents run in isolated virtual machines, so they continue when your laptop is closed. Since August 2026 they can hold a long-running objective with the /goal command and watch a pull request, a Slack thread or a schedule. Cursor Projects, in beta since 10 September 2026, goes further: a coordinator agent plans a larger piece of work such as a migration, hands parts of it to other agents and keeps shared notes so later agents reuse what earlier ones learned. Since 27 August you can also start a cloud agent from scratch without connecting a repository at all, which is handy for a prototype.
Review and release on team plans
Bugbot reviews pull requests on Premium and above. On Teams and Enterprise, Security Review (23 September 2026) posts one comment per pull request listing exploitable bugs, and Rollouts watches each change as it deploys and reports health per environment. Origin, Cursor's own code hosting, is in early beta on paid plans and can mirror GitHub with two-way sync of comments, reviews and merges.
Where Cursor falls short
- Your favourite model may now cost extra. If you chose Cursor because you liked Claude or GPT inside it, note that the new plans do not include any allowance for them. Premium and above can use them at API list price on top of the subscription, which makes monthly costs harder to predict. Set a spending cap.
- Plus is limited. Plus gives you the Agent with Cursor's own models at fixed settings (Fast off, medium effort, the 256k context window for Grok 4.7) and no Auto mode, cloud agents, Automations, Bugbot, Origin or your own API keys. For serious coding, Cursor itself recommends Premium.
- Agents still make confident mistakes. An agent that edits twenty files can introduce twenty subtle problems. Keep changes small enough to review, and keep your test suite meaningful.
- Extensions are a real risk. On 29 October 2026 Cursor reported that malicious versions 2.1.312 and 2.1.313 of the Anthropic.claude-code extension were available in its VS Code Extension Marketplace for several hours and tried to install an SSH backdoor. Cursor removed them and published remediation steps. Install only extensions you need, and check the publisher.
- It is heavy for non-developers. If you want a quick script or a one-off formula, a general assistant is simpler. Cursor pays off when you have a codebase.
How to get better results from Cursor
- Plan before you build. Press Shift+Tab in the chat input to switch to Plan Mode. The agent researches the code, asks clarifying questions and writes a plan you can edit before any code changes. Cursor also suggests Plan Mode on its own for tasks that sound complex.
- Write your rules once. Put project conventions in .cursor/rules as .mdc files, or in a plain AGENTS.md file at the root of the repository. Rules are added to the start of the model's context, so the agent follows your style, folder structure and testing habits every time. A plain .md file inside .cursor/rules is ignored, so use the .mdc extension there.
- Point at the right files. Mention the files, folders or rules that matter with @ in the chat. A smaller, relevant context beats a vague request over the whole repository.
- Ask for tests first. Tell the agent to write or run a failing test that shows the bug, then fix it. That gives you a clear pass or fail instead of a change that only looks right.
- Isolate experiments. Use /worktree for anything risky, and /best-of-n when you want to compare two or three models on the same task.
- Review in small pieces. Ask for one logical change per run and read the diff before accepting it. Long unattended runs are fine for cloud agents, but merge only what you have read.
Our Cursor lessons walk through Plan Mode and rules with a short exercise each, and our stacks show where Cursor fits next to a general assistant.
Privacy settings: where Privacy Mode lives and what it does
Privacy Mode is the setting that matters. When it is on, Cursor says your code is not used for training by Cursor or by model providers, and it has zero data retention agreements with its providers. When it is off, Cursor's data-use page says it may use and store codebase data, prompts, editor actions and code snippets to improve its AI features and train its models. Switch it on in Cursor's settings. On Teams, an admin can enable Privacy Mode for the whole team from the team dashboard under Settings and enforce it so members cannot turn it off. It is on by default for Enterprise.
Three details are easy to miss. Requests still pass through Cursor's servers even if you bring your own API key, because that is where Cursor builds the final prompt. Cloud agents are the one feature that stores your code, as encrypted copies kept while the agent runs and deleted afterwards, so skip them if your policy forbids code storage. And a few models fall outside the zero retention agreements: Cursor's documentation names Claude Fable 5.1 and Claude Fable 5, whose inputs and outputs Anthropic keeps for safety review, so under Privacy Mode they need admin approval before use.
Which Cursor plan suits whom
| Plan | Good for | Watch out for |
|---|---|---|
| Free | Trying the Agent on a side project | Limited requests, no third-party models |
| Plus | Light use with Cursor's own models | No cloud agents, Bugbot, Auto or Claude and GPT |
| Premium | Most working developers; Cursor's pricing page recommends it for new AI users | Third-party models billed on top |
| Super | Daily Agent users, 15x Plus usage | Only worth it if you hit Premium's weekly limit |
| Ultra | Power users running several agents or Automations, 50x Plus usage | Expensive if your usage is uneven |
| Teams and Enterprise | Companies that need SSO, enforced Privacy Mode, audit logs and access controls | Enterprise is custom priced |
Pick Premium if you code for a living and want Cursor's agents, cloud agents and Bugbot. Stay on a legacy plan for now if your workflow depends on included Claude or GPT usage, because switching ends that allowance. Choose Teams rather than personal plans the moment you work on company code, because enforced Privacy Mode is the only way to be sure nobody on the team leaves it off. Check Cursor's pricing page for current prices, since this is the second pricing change in a year.
Cursor or something else?
Cursor is not the only AI coding tool, and the new model rules may change your answer. If you mainly want Claude's models and a terminal workflow, read Claude vs Cursor. If your company already pays for GitHub, compare it with GitHub Copilot. For the wider field, see Cursor alternatives and the best AI for coding, or line tools up yourself in our comparison tool.
Prompts to copy
Map an unfamiliar codebase
I am new to this repository. Read the project structure and the main entry points, then explain in plain English what the app does, how a request flows from [entry point] to the database, and the five files I should read first. Do not change any code.
You get: A guided tour of the code with a reading list
Plan a feature before building it
Plan, do not code yet: I want to add [feature] so that [user outcome]. List the files you would change, the order of the changes, the tests you would add and any decisions you need from me. Ask me questions where the requirements are unclear.
You get: An editable plan and the questions you need to answer
Fix a bug with a failing test first
Here is the bug: [describe behaviour and steps to reproduce]. First write a test that fails because of this bug and run it. Then fix the cause, not the symptom, run the full test suite and summarise what you changed and why.
You get: A proven fix with a test that guards against regressions
Write project rules from your code
Study this codebase and draft an AGENTS.md file that captures our conventions: folder structure, naming, error handling, testing approach and anything an agent should never do. Keep it under 60 lines and quote real examples from the code.
You get: A rules file the agent will follow on every run
Refactor safely in small steps
Refactor [module or function] to [goal, such as remove duplication or split into smaller functions] without changing behaviour. Work in steps of no more than three files, run the tests after each step and stop if any test fails.
You get: A refactor you can review step by step
Review your own changes before a pull request
Review the changes on this branch compared with main as a strict senior reviewer. List bugs, security issues, missing tests and unclear names, ordered by severity, with the file and line for each. Do not fix anything yet.
You get: A prioritised review checklist before teammates see it
Explain an error you do not understand
I get this error when I run [command]: [paste error]. Find where it comes from in our code, explain the cause in two or three sentences a junior developer would understand, and propose the smallest fix. Show me the diff before applying it.
You get: A plain explanation and a minimal fix
Upgrade a dependency without surprises
We want to upgrade [library] from [current version] to [target version]. Find every place we use it, list the breaking changes that affect us, then make the changes in a separate worktree and run the tests. Report anything you could not resolve.
You get: An isolated upgrade with a list of open issues