Building an MCP Server

The Model Context Protocol (MCP) is an open standard for exposing tools, resources, and prompts to any compatible LLM client. Write a server once; every MCP-aware app can use it. This is the deep-dive companion to Building MCP Tools for a Local LLM.

#The three primitives

#A minimal tool server (Python)

from mcp.server.fastmcp import FastMCP
mcp = FastMCP("demo")

@mcp.tool()
def add(a: int, b: int) -> int:
    """Add two numbers."""
    return a + b

if __name__ == "__main__":
    mcp.run()   # stdio transport by default

#What makes a good MCP server

• Clear tool names + typed args + docstrings — the model picks tools from these (Agentic Tool-Use Loops).
• Small, composable tools beat one giant do-everything tool.
• Return errors as data so the model can recover.
• Choose transport: stdio (local subprocess) or HTTP+SSE (networked).

#Running it reliably

Run servers as supervised services that auto-restart and survive logout — see Building MCP Servers as systemd User Services.

Related: Building MCP Tools for a Local LLM · Building MCP Servers as systemd User Services · Agentic Tool-Use Loops