Multi-server MCP setup

You’ll connect one agent to several MCP servers at once (a local git server, a filesystem server and a hosted HTTP server) through a single MCPRegistry. Use this pattern when the tools an agent needs are spread across MCP servers; the registry connects them all, merges their tool lists, and routes each call to the right server.

from loomflow import Agent
from loomflow.mcp import MCPRegistry, MCPServerSpec
 
registry = MCPRegistry([
    MCPServerSpec.stdio(
        name="git",
        command="uvx",
        args=["mcp-server-git", "--repository", "/Users/me/code/myrepo"],
    ),
    MCPServerSpec.stdio(
        name="fs",
        command="npx",
        args=["-y", "@modelcontextprotocol/server-filesystem", "/Users/me/workspace"],
    ),
    MCPServerSpec.http(
        name="hosted",
        url="https://example.com/mcp/",
        headers={"Authorization": "Bearer ..."},
    ),
])
 
agent = Agent(
    "You are a developer assistant.",
    model="claude-opus-4-7",
    tools=registry,
)

Tool names that are unique across servers keep their bare name. When two servers expose the same name, both are registered under a qualified server__tool name: if git and fs both had a status tool, the agent would see git__status and fs__status (a double underscore, because model providers reject . in tool names). The qualified form is always accepted at call time, as is the dotted git.status form earlier releases used, and the bare form is accepted while it is unambiguous. The registry strips the prefix before forwarding the call to the owning server.

To mix MCP tools with your own @tool functions, wrap both in an ExtendedToolHost, as in the research agent recipe.