cursor-sdk-mcp

An MCP (Model Context Protocol) server that lets MCP clients such as Claude Code invoke the Cursor SDK and use Cursor Agent capabilities as tools for local and cloud coding workflows.

What it does

The server speaks MCP over stdio and exposes these tools:

Under the hood it uses @cursor/sdk's local and cloud runtimes. Local agents read/write files on disk from this Node process. Cloud agents run in Cursor-hosted or self-hosted environments and can use Cursor account/team MCP configuration from cursor.com/agents.

Related MCP server: cursor-agent-bridge

Requirements

• Node.js >= 22.13 (required by @cursor/sdk).

• A Cursor API key. Create one at the Cursor Dashboard → API Keys (user key) or Team settings (service account key).

Install & build

npm install
npm run build

Configure

Set your API key (see .env.example):

export CURSOR_API_KEY="your-cursor-api-key"
# Optional: default model id used when a tool call omits one (default: "auto")
export CURSOR_MCP_DEFAULT_MODEL="auto"

Use with Claude Code

Register the server with Claude Code (stdio):

claude mcp add cursor-sdk -- node /absolute/path/to/cursor-sdk-mcp/dist/index.js

Or add it to your MCP client config manually:

{
  "mcpServers": {
    "cursor-sdk": {
      "command": "node",
      "args": ["/absolute/path/to/cursor-sdk-mcp/dist/index.js"],
      "env": { "CURSOR_API_KEY": "your-cursor-api-key" }
    }
  }
}

Then ask Claude Code to, for example, "use the cursor agent to refactor src/auth.ts", and it will call cursor_run_local_agent or the compatibility cursor_run_agent.

Local agent usage

Use cursor_run_local_agent when the agent should work in local directories:

{
  "prompt": "Refactor src/server.ts to split schemas into src/schemas.ts and run tests.",
  "cwd": "/absolute/path/to/repo",
  "model": "auto",
  "mode": "agent",
  "settingSources": ["project", "user", "plugins"],
  "autoReview": true
}

settingSources controls which Cursor MCP/settings layers the local runtime loads from disk. You can also pass inline mcpServers, agents, and sandboxOptions for one-off tool wiring and safety. For local cursor_follow_up, lifecycle, and artifact calls, pass the same cwd used to create the agent so the SDK can find persisted local agent state.

Live progress and cancellation

cursor_run_agent, cursor_run_local_agent, cursor_run_cloud_agent, and cursor_follow_up run to completion in a single tool call. Because a real coding run can take minutes, the server keeps the call responsive:

• Progress streaming. When the MCP client sends a progressToken with the request (most clients do this automatically when you register a progress callback), the server streams each agent step — assistant text, tool calls, status changes — back as notifications/progress. Clients that reset their request timeout on progress (set resetTimeoutOnProgress) won't time out on long runs, and the user sees what the agent is doing instead of a silent wait.

• Cancellation. If the client cancels the tool call (its AbortSignal fires), the server cancels the underlying Cursor run, so you stop paying for work you no longer need.

Progress streaming is best-effort: if a client doesn't request progress, or the runtime doesn't support streaming, the tool still returns the same final result.

Cloud agent usage

Use cursor_run_cloud_agent when the agent should run in Cursor Cloud:

{
  "prompt": "Add tests for the auth middleware and open a PR.",
  "repos": [{ "url": "https://github.com/your-org/your-repo", "startingRef": "main" }],
  "model": "auto",
  "autoCreatePR": true
}

repos can be omitted for an empty workspace or a named cloud environment. Cloud agents can use inline MCP plus Cursor account/team MCP configured at cursor.com/agents. Use the lifecycle tools to inspect runs, cancel work, and fetch artifacts.

Cursor MCP and skills strategy

• Local agents use inline MCP unless settingSources includes project, user, or plugin settings.

• Local stdio MCP servers may include cwd; cloud stdio MCP servers must not include cwd.

• Cloud agents use inline MCP plus user/team MCP from cursor.com/agents.

• OAuth MCP must already be authorized in Cursor before local reuse.

• Cursor skills can live in .cursor/skills/, .agents/skills/, ~/.cursor/skills/, or ~/.agents/skills/.

• Claude Code skills are separate from this integration and are not loaded by Cursor Agent through this MCP server.

Development

npm run dev        # run from source with hot reload (tsx)
npm run typecheck  # type-check only
npm run lint       # eslint
npm test           # integration tests (fake Cursor backend, no network)
npm run demo       # spawn the built server over stdio and call its tools
RUN_AGENT=1 npm run demo   # additionally run a real Cursor Agent (needs CURSOR_API_KEY)

How it's structured

• src/cursor.ts — CursorService interface + CursorSdkService (the @cursor/sdk wrapper).

• src/server.ts — builds the McpServer and registers the tools. Decoupled from the SDK via CursorService so tests can inject a fake backend.

• src/index.ts — entry point; wires the real service to a StdioServerTransport.

• tests/server.test.ts — connects an in-memory MCP client to the server and exercises every tool.

• examples/demo-client.ts — end-to-end demo over a real stdio transport.

License

MIT