🚀 AnyCowork is in active development - Star us on GitHub! 🚀
Documentation
Model Context Protocol

Model Context Protocol (MCP)

AnyCowork is built on the Model Context Protocol (MCP), an open standard that enables AI assistants to connect to data and tools in a standardized way.

What is MCP?

MCP (opens in a new tab) is a protocol that standardizes how AI models interact with external context. Instead of building custom integrations for every database, file system, or API, MCP provides a universal language for these connections.

Think of it like a USB port for AI applications. Just as USB let you connect any device to any computer, MCP lets you connect any data source to any AI assistant.

Why AnyCowork Uses MCP

We chose MCP as the backbone of AnyCowork for three reasons:

  1. Interoperability: You can use existing MCP servers from the community (e.g., for PostgreSQL, Google Drive, GitHub) without waiting for us to build an integration.
  2. Security: MCP allows for granular permission control. You decide exactly what resources an agent can access.
  3. Future-Proofing: As the ecosystem grows, AnyCowork grows with it.

How It Works in AnyCowork

AnyCowork acts as an MCP Client. It can connect to local or remote MCP Servers.

Architecture

Supported MCP Features

  • Resources: Agents can read data from external sources (e.g., reading a file, fetching a database row).
  • Tools: Agents can execute commands (e.g., running a SQL query, committing code).
  • Prompts: Reusable prompt templates provided by servers.

Getting Started with MCP

(Note: MCP implementation is currently in active development. The following features are part of the v0.2.0 roadmap.)

connecting a Server

  1. Open Settings -> Integrations.
  2. Click Add MCP Server.
  3. Choose the transport type (Stdio or SSE).
  4. Enter the command to run the server (e.g., npx -y @modelcontextprotocol/server-postgres).

Building Your Own Server

You can build custom MCP servers to expose your internal tools to AnyCowork.

# Example: A simple MCP server in Python
from mcp.server import MultiServer
 
async def main():
    server = MultiServer("my-tools")
    
    @server.tool("hello")
    async def say_hello(name: str) -> str:
        return f"Hello, {name}!"
        
    await server.serve_stdio()

Recommended MCP Servers

Skills & SandboxSystem Architecture