Using MCP Servers - Marvin

Pass an MCP connection through Agent(mcp_servers=[...]). Pydantic AI provides the client integration.

Pydantic AI 2.x

Use MCPToolset for an HTTP MCP server:

import marvin
from pydantic_ai.mcp import MCPToolset

server = MCPToolset("http://localhost:8000/mcp")
agent = marvin.Agent(mcp_servers=[server])
answer = agent.run("List the tools you can use")

The server must already be running. To connect a subprocess, pass a FastMCP StdioTransport to MCPToolset; see Pydantic AI’s MCP client documentation.

In-process FastMCP

Install the optional server package:

uv add "marvin[mcp]"

import marvin
from fastmcp import FastMCP

server = FastMCP("calculator")

@server.tool
async def add(a: int, b: int) -> int:
    return a + b

agent = marvin.Agent(mcp_servers=[server])
answer = agent.run("Use add to calculate 23 + 19", result_type=int)

Pydantic AI 1.x

Existing MCPServerStdio and HTTP server instances remain supported when using Pydantic AI 1.x. Those classes were replaced by MCPToolset in 2.x; use the API that matches your installed Pydantic AI version. A Thread context keeps managed server connections available across calls and closes them on exit. See threads.

⌘I