cursor-agent-bridge

An MCP server that lets Claude Code (or any MCP client) hand off coding tasks to the local Cursor Agent CLI, running in a specific project workspace. Each workspace gets its own persistent Cursor session that automatically resumes across calls, so Cursor remembers prior turns for that project even after you've worked on other projects in between.

Useful for running Claude and Cursor side by side on the same codebase — Claude plans and reviews, Cursor executes, and both stay in sync per project.

Setup

{
  "mcpServers": {
    "cursor-agent-bridge": {
      "command": "node",
      "args": ["/path/to/cursor-agent-bridge/index.js"]
    }
  }
}

Requires the cursor-agent CLI installed and on PATH (or set CURSOR_AGENT_PATH).

Related MCP server: cursor-mcp-server

Tools

• cursor_agent_delegate — send a prompt + workspace path to Cursor Agent, resuming that workspace's session if one exists.

• cursor_agent_sessions — list all workspaces with an active/remembered session and when they were last used.

How it works

This bridge delegates project-scoped coding work to the local Cursor Agent CLI.

Shell and SSH permission

File editing is the default mode. Unrestricted shell commands, SSH, and network access are disabled unless the caller opts in for an individual tool call.

A calling Claude or GPT agent must:

1. Explain that enabling shell mode lets Cursor run local and remote commands without further approval.

2. Ask the user for explicit permission in the current conversation and wait.

3. Set both allow_shell: true and shell_permission_confirmed: true only after approval.

Permission is per call and must not be inferred from an earlier task. When enabled, the bridge launches Cursor with --force --sandbox disabled, allowing it to use the Mac's SSH configuration, credentials, and Tailscale routes.

Clients that already loaded the MCP server must restart or reconnect it to see schema or implementation changes.

Timeout

Each call defaults to a 10-minute timeout. Any calling AI (Claude, ChatGPT, etc.) can raise this per call via the timeout_ms parameter, up to a hard cap of 60 minutes.