MCP Servers¶
Connect an external Model Context Protocol server and expose its tools to an agent.
Minimal walkthrough¶
Install the extra, then connect a server through the MCP manager — its remote tools become ordinary Tools in your pool:
from neurosurfer.mcp import McpManager
from neurosurfer.tools import default_pool
manager = McpManager()
await manager.connect("filesystem", command="npx", args=["-y", "@modelcontextprotocol/server-filesystem", "/data"])
pool = default_pool()
pool.extend(manager.tools()) # MCP tools now callable by the agent
agent = AgenticLoop(provider=provider, tools=pool, system_prompt="…",
guardrails=Guardrails(), io=AutoIO(), cwd=Path.cwd())
MCP tool calls are gated by mcp_policy in Guardrails, like any other action.
Full notebook¶
The Colab notebook walks through connecting a server, listing its tools, and persisting server configs.
Next: MCP guide · Tutorial 5 →