MCP

Plug into the MCP ecosystem.

Model Context Protocol is becoming the open standard for AI agent tools. Agentmatic ships a first-class MCP client and server — no wrappers, no glue code.

Connect to an MCP server

from agentmatic.tools.mcp import MCPClient
from agentmatic.prebuilt import create_react_agent

# Stdio transport — spawns the server as a subprocess.
mcp = MCPClient.stdio(["npx", "-y", "@modelcontextprotocol/server-filesystem", "/tmp"])
tools = await mcp.list_tools_as_agentmatic()  # filesystem.read, filesystem.write, ...

agent = create_react_agent(llm=OpenAI(), tools=tools)
print(await agent.invoke("Summarize the README at /tmp/project/README.md"))

Expose your tools as an MCP server

from agentmatic.tools.mcp import serve_mcp
from my_tools import calculator, db_query

# Publish any Agentmatic tool registry as an MCP server over stdio.
if __name__ == "__main__":
    serve_mcp([calculator, db_query], transport="stdio")

Why MCP matters

• Process isolation. Tools that touch the filesystem or shell run out-of-process — safer than embedding them in your Python.
• Language-agnostic. An MCP server in Rust, Go, or TypeScript is callable from any Agentmatic SDK.
• Ecosystem. Filesystem, git, browser (Playwright), Postgres, GitHub, Slack, Gmail — there's an MCP server for it.
• Portability. The same tool you ship to your Agentmatic agent is reusable by Claude Desktop, Cursor, Cline, and anything else MCP-compatible.

Server discovery

Agentmatic respects mcp.json the same way Claude Desktop and Cursor do — drop a config and your agent finds the same servers your IDE uses.

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/me/projects"]
    },
    "git": { "command": "uvx", "args": ["mcp-server-git"] }
  }
}