pi-mcp-server

MCP server that wraps the Pi coding agent as tools. Use Pi as a sub-agent from Claude Desktop, Cursor, or any MCP client.

Why use this?

Pi is a fast, lightweight coding agent. This server exposes it via MCP so your primary AI can delegate coding tasks to Pi — getting a second opinion, running in parallel, or using a different model.

Installation

npm install -g pi-mcp-server

Or run directly with npx:

npx pi-mcp-server

Quick Start

Claude Desktop

Add to your Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "pi": {
      "command": "npx",
      "args": ["-y", "pi-mcp-server"],
      "env": {
        "ANTHROPIC_API_KEY": "sk-ant-..."
      }
    }
  }
}

Cursor

Add to .cursor/mcp.json in your project:

{
  "mcpServers": {
    "pi": {
      "command": "npx",
      "args": ["-y", "pi-mcp-server"],
      "env": {
        "ANTHROPIC_API_KEY": "sk-ant-..."
      }
    }
  }
}

Environment Variables

Tools

pi

Start a new Pi coding agent session.

Input:

• prompt (string, required) — The task or question
• provider (string, optional) — LLM provider override
• model (string, optional) — Model ID override
• thinkingLevel (string, optional) — Thinking level override
• cwd (string, optional) — Working directory override

Output: Response text prefixed with [thread_id: <uuid>] for follow-up.

pi-reply

Continue an existing session.

Input:

• threadId (string, required) — Thread ID from a previous pi call
• prompt (string, required) — Follow-up message

Output: Response text from the agent.

Sandbox Mode

Enable sandbox mode for isolated command execution:

PI_MCP_SANDBOX=true npx pi-mcp-server

Requires the optional dependency @anthropic-ai/sandbox-runtime:

npm install @anthropic-ai/sandbox-runtime

Development

git clone https://github.com/pandysp/pi-mcp-server.git
cd pi-mcp-server
npm install
npm run build
npm test

# Smoke test (requires API key)
ANTHROPIC_API_KEY=sk-ant-... node test/smoke.mjs

License

MIT