What Is Cursor AI? The Editor, Its Agent and Who It Suits

Cursor AI is a code editor with an AI agent built into it. It is based on the VS Code codebase, so it looks and behaves like VS Code, opens the same projects and imports your settings, but the center of the window is an agent that reads your code, edits files, runs terminal commands and checks its own work. Around that sit Tab, an autocomplete that predicts your next edit; a choice of models from several labs; rules that tell the agent how your project works; and MCP servers that connect it to other tools. The same agent also runs in a command line, in the cloud and on iPhone. It suits people who want to watch changes land in the file they are reading.

The editor: VS Code underneath

Cursor’s VS Code migration guide↗ (opens in a new tab) says it is based upon the VS Code codebase and is regularly rebased onto recent VS Code versions, often slightly older ones for stability. For you that means:

• One-click import of your VS Code extensions, themes, settings and keybindings, under Cursor Settings, General, Account.
• The same default keyboard shortcuts as VS Code.
• Extensions come from the Open VSX registry, not the VS Code Marketplace, so most popular extensions are there but not every one.
• Installers for macOS 12 and later, Windows 10 and later, and Linux (apt and dnf packages, or an AppImage). Cursor and VS Code can sit side by side on the same machine.

On top of the editor, Tab↗ (opens in a new tab) is Cursor’s autocomplete. It suggests multi-line edits from your recent changes, the code around the cursor and linter errors, and after you accept one, pressing Tab again jumps to where it predicts you will edit next, including in another file.

The agent

Press Cmd+I (Ctrl+I on Windows and Linux) and the agent opens in a side pane. Cursor’s agent overview↗ (opens in a new tab) describes it as three parts: instructions (a system prompt plus your rules), tools, and the model you pick. The tools cover searching files and the web, reading files and images, editing, running shell commands, controlling a browser to check visual changes, and asking you clarifying questions while it keeps working. There is no limit on the number of tool calls in a task.

• Modes: Agent edits, Ask only reads and answers, Plan writes a plan for you to approve first, and Debug chases a bug with logging. Which to use when is covered in Cursor agent vs Ask vs Plan.
• Checkpoints: taken automatically before significant changes. Restoring one reverts files only, not the conversation, and not anything a command did outside your files. They are stored locally and are no substitute for Git.
• Steering: queue a message for when the agent finishes, or send one now and it is picked up at the agent’s next tool call.
• Projects and goals: for bigger work, a Project has a coordinator agent plan and delegate to other agents, and /goal gives the agent a long-lived objective (it is still rolling out).

The Agents Window

Since Cursor 3, released on April 2, 2026, Cursor has had a second interface. The Agents Window↗ (opens in a new tab) is an agent-first workspace across repositories and environments: local, cloud and remote SSH. It adds parallel agents, worktrees so each task has its own checkout, a diff view for committing and managing pull requests, and handoff between local and cloud. Cursor says the editor is the better choice when you want the classic IDE with VS Code extensions and split panes, and the Agents Window when agents write most of your code. You can switch between them at any time with Cmd+Shift+P.

Models

Cursor runs frontier models from OpenAI, Anthropic, Google and others, alongside its own Composer model, and tunes the agent’s instructions and tools per model. You pick one per chat, or choose Auto and let Cursor pick. On Teams and Enterprise plans, Auto is backed by Cursor Router, which you steer with an optimization mode: Cost, Balance or Intelligence. Usage is split into two pools, one for Cursor’s own group of models and one for third-party models at their API rates, so the model you choose affects how fast you use up what your plan includes. Model lists change often; check the Models page in Cursor’s docs rather than any list in a blog post.

Rules

Rules are standing instructions the agent reads at the start of its context. Cursor’s rules documentation↗ (opens in a new tab) lists four kinds:

• Project rules: .mdc files in .cursor/rules, committed with the code. Frontmatter decides when each applies: always, when the agent judges it relevant from its description, when a matching file is in context, or only when you @-mention it. A plain .md file there is ignored.
• User rules: yours, in every project.
• Team rules: managed from the dashboard on Teams and Enterprise plans.
• AGENTS.md: plain Markdown with no frontmatter, at the root or in any subdirectory, where more specific files take precedence.

Put shared conventions in AGENTS.md, which other tools read too, and keep .cursor/rules for what only Cursor needs. Ready-made files are in Cursor rules examples, and which tool reads which file is in AI context files compared.

MCP

Cursor is an MCP client. Its MCP documentation↗ (opens in a new tab) lists three transports (stdio for local servers, and SSE or Streamable HTTP for remote ones, which can sign in with OAuth) and support for tools, prompts, resources, roots, elicitation and the MCP Apps extension. Install servers from the Customize page or write them into .cursor/mcp.json in a project, or ~/.cursor/mcp.json for every project. By default Cursor asks before it uses an MCP tool, and on Enterprise an admin can allowlist which servers and tools may run.

{
  "mcpServers": {
    "fenbs": {
      "url": "https://fenbs.ai/api/mcp"
    }
  }
}

Beyond the editor: CLI, cloud agents, iOS and JetBrains

• The CLI: the same agent in a terminal, started with agent, using the same rules and MCP configuration. Setup and flags are in Cursor CLI.
• Cloud agents: the agent in an isolated virtual machine that clones your repository, works on a branch and pushes it, started from the editor, the web, Slack, Linear or a GitHub comment. See Cursor cloud agents.
• Cursor for iOS: a native app for starting agents, following them and reviewing and merging their pull requests from an iPhone or iPad. Android is planned.
• JetBrains IDEs: on a paid plan, Cursor’s agent can run inside IntelliJ IDEA, PyCharm and the rest through the Agent Client Protocol and JetBrains AI Assistant.

Plans, by name

As of September 30, 2026, Cursor’s individual plans are Hobby, which is free with limited agent use, Pro, Pro+ (written Pro Plus in some docs) and Ultra; for companies, Teams and Enterprise. A lower-cost Start plan exists for developers in India only. Paid plans add more agent usage, access to third-party models, cloud agents and Bugbot, its pull request reviewer; team plans add central billing, team rules, SSO and usage analytics. Check Cursor’s current plans for limits, which change often.

Who Cursor suits, and who a terminal agent suits

• Cursor: you start from an open file, like to see each diff land in the editor, and want autocomplete and an agent in one window. Coming from VS Code, the move is nearly free.
• A terminal agent such as Claude Code: you start from a prompt, script your tools, run the same agent in CI, or work over SSH. It sits beside whatever editor you already use. The two are compared in Claude Code vs Cursor.
• Both: plenty of people run a terminal agent in Cursor’s integrated terminal. Share rules through AGENTS.md and give each agent its own branch.
• Neither: if you want to stay in plain VS Code or a JetBrains IDE, or want a different workflow, see Cursor alternatives. GitHub’s own desktop app for running agents is covered in the GitHub Copilot app.

Where the plan lives when the chat ends

Cursor’s plans and chats belong to one person’s editor, and a cloud agent starts from the repository alone. fenbs keeps the work list where every agent and every teammate can see it: tasks in To Do, Next Up, In Progress and Completed, each with a note for the problem, a plan for how it will be done, and a test status. With the mcp.json above, Cursor signs in with OAuth in your browser, or you can issue a token under Settings with a name, scopes and an optional expiry. Add a line to AGENTS.md: read the task with fenbs_get_item before starting, write an approved plan into it with fenbs_update_item, and comment with what changed when you stop. History then shows each change under the assistant’s name. fenbs has no sprints, due dates or WIP limits; it is a list, not a process.