cursor-delegate-mcp

Cursor Delegate is an MCP (Model Context Protocol)-based bridging tool used to delegate multi-file code implementation tasks from your commonly used frontier coding agents (such as Claude Code, ChatGPT/Codex, GitHub Copilot CLI, or other MCP clients) to Cursor's CLI agent (Composer 2.5). Its design goal is to let your best coding agent focus on the parts where judgment is most critical: understanding the task, formulating a plan, and reviewing the results, while handing repetitive code writing to Cursor.

Main capabilities:

• Multi-file implementation delegation: the entire delegation process requires only one call—input a brief, return a result. Composer 2.5 is designed for throughput on multi-file edits, so you can review a completed diff without watching edits land file by file.
• Preserve frontier quality: your assistant is responsible for understanding the task, writing a precise brief, and reviewing the completed diff; Composer 2.5 runs the implementation loop under that brief. The bundled skill incorporates the review step into the loop, and the agent responsible for scoping the work reads the changed files before reporting.
• Separate quota: Composer and Grok consume from the Cursor Models usage pool, separate from third-party model usage on supported Cursor plans. A Claude or Codex subscription is used for writing briefs and reviewing, while Cursor handles implementation. If you pay for API usage by token, the implementation portion shifts to the Cursor capacity you already have.
• Native plugins: can be installed into Claude Code, ChatGPT/Codex, or GitHub Copilot CLI, and then you only need to say "delegate this to Cursor". The shared skill teaches your agent how to delegate correctly.
• No stuck runs: if Cursor needs clarification, it ends the current turn and returns the question as a normal result. Your agent reads the question and answers by resuming the same session—no modal window blocks waiting for input.
• Clean, named results: the final answer, changed files, session ID, and plan are returned as a compact JSON payload, with no scraping or parsing needed.
• Plan first (plan mode): Cursor drafts a plan, you review it, and then the same session implements that plan.
• Q&A mode (ask mode): ask and answer questions about the codebase.
• Self-diagnosis: the doctor tool tells you exactly what is missing when the configuration is incorrect.
• Portable: pure stdio MCP, can run in VS Code, JetBrains, Windsurf, Visual Studio, OpenCode, Antigravity, and other environments.

How to use:

• Requires Node.js 20+ and a logged-in Cursor CLI (cursor-agent login).
• Configure the server in an MCP client with the command npx and the argument -y cursor-delegate-mcp.
• The configuration location differs by client: VS Code adds the server under Settings → Tools → AI Assistant → Model Context Protocol (MCP); JetBrains uses the same mcpServers block as the generic snippet; Windsurf uses the same mcpServers block (Cascade supports up to 100 tools for all servers); Visual Studio uses servers as the top-level key (note that it is not mcpServers), requires Visual Studio 2026 or 2022 17.14+, and will also discover .mcp.json next to the solution as well as .vscode/mcp.json; OpenCode does not use mcpServers, local servers go under mcp, with type as "local" and command as an array; Antigravity uses the same mcpServers block, configured in the agent panel under MCP Servers → Manage MCP Servers → View raw config, Antigravity 2.0, the IDE, and the CLI share the Gemini configuration file, and the first run may require approval of server tools; Zed uses the same mcpServers block, with workspace files taking precedence; in the VS Code extension, add it through Settings → MCP → Add Server → Local (stdio); on Windows, if npx cannot be found, you can use ccommand cmd and parameters /c, npx, -y, cursor-delegate-mcp.
• You can also add it manually, or run Chat: Install Plugin From Source in Claude Code and pass in the repository URL to install the plugin.
• plan and ask are instructions that direct the agent, and workspace is its working directory.

Example usage: "Delegate to Cursor: migrate src/api from callbacks to async/await and update the tests, then walk me through what changed." The whole loop is: Claude writes the brief, Cursor handles the files, and Claude walks you through reviewing the diff.

Use cases:

• Developers who want to hand off multi-file boilerplate implementation to Cursor while using Claude or Codex for task understanding and result review.
• Users who want to avoid consuming frontier model quota on repetitive coding.
• Developers using an editor or agent client that supports MCP and has the Cursor CLI configured.

Limitations:

• Requires Node.js 20+ and a logged-in Cursor CLI.
• It runs based on MCP, so the client needs to support MCP (some clients have a tool count limit, such as Cascade with a maximum of 100 tools).
• Licensed under MIT.
• Regarding specific charges, pricing, platform support details, and results, the materials do not provide complete information, so it is not provided for now.