How to Build a Custom Python MCP Tool for OpenClaw

If you're using OpenClaw, you've probably hit a wall where your local AI agent needs to access a specific internal database, an API, or a custom script. The solution is the Model Context Protocol (MCP). In this tutorial, I'll show you exactly how to write a simple Python MCP server and connect it to OpenClaw.

What is MCP in OpenClaw?

The Model Context Protocol (MCP) is a standardized way for AI models to interact with local tools. Instead of hardcoding API keys into your agent, you expose an MCP server over stdio or HTTP, and OpenClaw automatically discovers the available tools.

Step 1: Writing the Python Tool

I recently needed an agent to check server logs. Here's the minimal Python MCP server I built using the official SDK:

from mcp.server.fastmcp import FastMCP

mcp = FastMCP("log_checker")

@mcp.tool()
def get_recent_logs(lines: int = 50) -> str:
    """Fetch the most recent server logs."""
    try:
        with open("/var/log/syslog", "r") as f:
            return "".join(f.readlines()[-lines:])
    except Exception as e:
        return f"Error reading logs: {e}"

if __name__ == "__main__":
    mcp.run()

Step 2: Configuring openclaw.json

Once the Python script is ready, you need to tell OpenClaw where to find it. Open your ~/.openclaw/openclaw.json and add the server configuration:

{
  "mcp": {
    "servers": {
      "log_checker": {
        "command": "uv",
        "args": ["run", "/path/to/log_checker.py"]
      }
    }
  }
}

Step 3: Testing the Agent

Restart your OpenClaw gateway, and run your agent. You can now ask: "Check the last 20 lines of the system logs." The agent will autonomously call your Python tool, read the output, and synthesize an answer.

Building custom tools is what makes local AI truly powerful. By isolating your logic in MCP servers, you keep your agents modular and secure.

Written by Matteo Giardino, a fractional CTO and developer who builds local AI tools and autonomous workflows.