MCP servers

pip install pycodeloop[mcp]

Connect to any Model Context Protocol server over stdio and expose its tools to the agent alongside the built-in ones.

As a library

from pycodeloop import CodeLoop, Config
from pycodeloop.core.mcp import MCPServer, load_mcp_tools
from pycodeloop.core.tools import DEFAULT_TOOLS
from pycodeloop.providers import GenericProvider

server = MCPServer(command="npx", args=["-y", "@modelcontextprotocol/server-filesystem", "."])
tools = DEFAULT_TOOLS + load_mcp_tools(server)

config = Config(provider=GenericProvider.from_json("templates/anthropic.json"), tools=tools)
flow = CodeLoop(config=config)

load_mcp_tools(server) connects, lists the server's tools, and returns each one wrapped as a regular Tool — Agent doesn't know or care that the call is going over stdio to a subprocess instead of running in-process.

From the CLI

One --mcp flag per server, command arg1 arg2 shell-quoted:

pycodeloop run "list every allowed directory" \
  --mcp "npx -y @modelcontextprotocol/server-filesystem ."

Lifecycle

MCPClient (pycodeloop.core.mcp.MCPClient) owns the server subprocess on a dedicated background event loop for the life of the process — MCP sessions are async and expect to live inside one async with block for their whole lifetime, so a background loop lets a synchronous Tool.run() call into a long-lived subprocess without blocking Agent.

Every MCP tool is dangerous = True by default (see Permission prompts) since a remote server's tools are opaque — its preview() renders the call as tool_name(arg=value, ...).