MCP

The Model Context Protocol  (MCP) is an open standard for exposing tools and data to AI applications, and many services and developer tools ship an MCP server. Loomflow speaks MCP natively: point an agent at stdio or HTTP MCP servers and their tools become the agent’s tools, handled like any other ToolHost.

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.http(
        name="hosted",
        url="https://example.com/mcp/",
        headers={"Authorization": "Bearer ..."},
    ),
])
 
agent = Agent(
    "You are a coding assistant.",
    model="claude-opus-4-7",
    tools=registry,
)

Tool name conflicts across servers are auto-disambiguated: git__commit and github__commit if both servers expose commit; just commit if only one does. The qualified form is always accepted at call time (so is the dotted git.commit from earlier releases), and the bare form while it’s unambiguous.

Read more

What MCP gets you

• Language-agnostic tools. MCP servers can be written in Python, Node, Rust, or any other language. The framework launches them as subprocesses (stdio) or talks to them over HTTP.
• Reusable tool servers. The MCP ecosystem already ships servers for git, filesystem, GitHub, Slack, browser, Figma. You don’t re-implement what already exists.
• The full protocol surface. Beyond tools, the registry aggregates resources and prompts across servers, refreshes tool lists automatically when a server announces listChanged, and can answer server-initiated sampling requests with a model you choose.
• Fault isolation. One dead server never takes down the others — its tools are skipped, its name lands in registry.unavailable, and it heals on the next refresh. See Server specs + registry.
• Auth and rate limiting on the server. Hosted MCP servers authenticate the framework via headers; the agent never sees the underlying credentials.

Install

pip install 'loomflow[mcp]'

Pulls in the official mcp SDK + httpx. Both SDK generations work: 1.x (from 1.8) and 2.x.

Compose with built-in tools

MCPRegistry is just a ToolHost. Mix MCP servers with regular @tool functions by wrapping the registry with ExtendedToolHost, which combines a base host with extra in-process tools:

from loomflow import Agent, tool
from loomflow.mcp import MCPRegistry, MCPServerSpec
from loomflow.architecture.tool_host_wrappers import ExtendedToolHost
 
@tool
async def custom_search(q: str) -> str:
    """My custom search tool."""
    ...
 
mcp_registry = MCPRegistry([
    MCPServerSpec.stdio("git", "uvx", ["mcp-server-git"]),
    MCPServerSpec.stdio("fs",  "npx", ["-y", "@modelcontextprotocol/server-filesystem", "."]),
])
 
# Combine the MCP registry with the custom tool into one host
agent = Agent(
    "...",
    model="claude-opus-4-7",
    tools=ExtendedToolHost(mcp_registry, [custom_search]),
)