The name is the joke: a cursor is the blinking line where you type. Cursor is a code editor, built from the same open foundation as Visual Studio Code, that puts an AI agent at the centre of the writing experience.
It is also the product most often confused with the models inside it. Cursor does not make Claude or Gemini or GPT. It is a workbench that can use them.
Checked on 30 September 2026 against the Cursor pricing and CLI pages and recent coverage of Cursor 3. Plans and included usage change often, so confirm on cursor.com/pricing.
What Cursor is made of
| Piece | What it does |
|---|---|
| The editor | A full code editor with tab completion that predicts your next edit, inline edits and chat about your code. |
| Agent mode | You describe a task. The agent plans, edits several files, runs terminal commands and tests, and proposes the result for your approval. |
| The Agents Window (Cursor 3, April 2026) | A workspace for launching and supervising several agents at once, some on your machine and some in the cloud, including ones started from Slack, GitHub, Linear or your phone. |
| Cloud agents | Agents that run on isolated virtual machines, so a long task keeps going after you close your laptop. Results come back for review. |
| Cursor CLI | The same agent in a terminal. |
| Models | A menu of frontier models from several companies, plus Cursor's own. |
It also supports the wider toolkit of modern agents: MCP servers, reusable skills and hooks.
Its own model: Composer
Besides routing to other companies' models, Cursor trains its own coding model, Composer. Coverage of Composer 2 (March 2026) describes it as purpose-built for agentic coding and built on an open base model with extra training. The point of a house model is speed and cost: for routine steps of an agent loop, a fast specialised model can be cheaper than calling the most expensive frontier model every time.
Agent mode, in plain words
Older assistants suggested code and waited for you to paste it. Agent mode closes the loop:
- You give a goal: "add a dark mode toggle and make the tests pass".
- The agent searches your codebase to understand it.
- It edits files, runs the build or tests, reads the errors and fixes them.
- You review a diff and accept or reject.
You stay the reviewer. The agent does the legwork. Good habits: keep tasks small and specific, say how to verify success ("run the test suite"), and read the diff before accepting.
The Cursor CLI
The Cursor CLI brings the agent to the terminal. Install it with:
curl https://cursor.com/install -fsS | bash
In the interactive view you can use / for commands, @ to attach files and ! to run shell commands. It also has a headless mode for scripts and automation, including GitHub Actions, so you can run the same agent in your build pipeline. It works alongside whichever editor you prefer.
How Cursor charges (and why it is different)
Cursor is a subscription with included usage, not a pure per-token service and not a flat "unlimited" plan. From the pricing page:
- Hobby: free, with limited agent requests.
- Pro: about $20 a month. Third-party coverage describes unlimited tab completions and a pool of included model usage.
- Pro+ (about $60) and Ultra (about $200): larger usage pools for daily and power users.
- Teams ($40 per user per month) and Enterprise (custom): shared billing, admin controls, single sign-on.
The key to understanding it is the usage model: each plan includes a set amount of model usage, priced according to which model you pick. A cheap fast model stretches your allowance a long way, and a top frontier model burns through it faster. When you use it all up, extra usage is billed after the fact, at the model's rates. So Cursor behaves like a prepaid allowance with overflow, sitting between a flat subscription and an API bill.
Cursor versus the model makers' own agents
| Cursor | Claude Code, Codex, Gemini CLI | |
|---|---|---|
| Made by | a separate editor company | the model makers |
| Models | several companies, plus Cursor's | one company's (Gemini CLI, Codex) or one family (Claude Code) |
| Interface | editor first, plus a CLI and cloud agents | terminal first, with editor extensions |
| Billing | one Cursor plan with pooled model usage | your plan with that company, or their API key |
Neither is better in general. If you want one polished place, a visual diff review and the freedom to swap models, Cursor is attractive. If you live in the terminal, want your model maker's newest features the day they ship, or want to pay a single vendor directly, the native agents fit better. Many developers use both, and some even run Claude Code inside Cursor's editor.
What you can do with Cursor
- Build a feature across many files in one go.
- Ask "how does this codebase work?" and get a guided tour.
- Fix bugs by pasting an error and letting the agent trace and patch it.
- Run several agents in parallel on separate branches.
- Run an agent in CI to review pull requests or triage issues.
Links
- Pricing: cursor.com/pricing
- CLI: cursor.com/cli
- Comparison with other agents: CLIs and agent mode