You understand MCP conceptually but don't know where to start? This tutorial builds two runnable servers (file reader + weather) from scratch in 10 minutes, and explains stdio vs SSE.

💡 What You Will Learn

You understand MCP conceptually but don't know where to start? This tutorial builds two runnable servers (file reader + weather) from scratch in 10 minutes, and explains stdio vs SSE.

📜 Table of Contents

• 1. Prerequisites
• 2. Install the SDK
• 3. First Server: File Reader
• 4. Run It and Connect a Client
• 5. Second Example: Weather Server
• 6. stdio vs SSE
• 7. One Server, Many Tools
• FAQ

MCP Server Tutorial 2026: Build Your First MCP Server from Scratch

MCP (Model Context Protocol) lets AI clients like Claude Desktop call external tools through a unified standard. Concepts are easy; getting started is the hard part. This hands-on tutorial builds two working servers in about 10 minutes.

1. Prerequisites

• Python 3.10+
• pip

No GPU or server needed.

2. Install the SDK

pip install mcp

3. First Server: File Reader

Register two tools so the AI can read your local files:

from mcp.server.fastmcp import FastMCP

mcp = FastMCP("file-reader")

@mcp.tool()
def read_file(path: str) -> str:
    """Read a local file"""
    with open(path, "r", encoding="utf-8") as f:
        return f.read()

@mcp.tool()
def list_files(directory: str) -> list[str]:
    """List files in a directory"""
    import os
    return os.listdir(directory)

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

Save as file_reader_server.py. The docstrings are the tool descriptions the AI uses to decide when to call each tool—write them clearly.

4. Run It and Connect a Client

python file_reader_server.py

The server communicates over stdio and stays silent until a client connects—that's normal. In Claude Desktop (or any MCP-capable client), add an MCP server with the command python /absolute/path/file_reader_server.py, then restart. Now you can say "read notes.txt on my desktop" and the AI will call read_file automatically.

5. Second Example: Weather Server

Tools can also call external APIs. Here's a weather tool using the free wttr.in service:

from mcp.server.fastmcp import FastMCP
import urllib.request

mcp = FastMCP("weather")

@mcp.tool()
def get_weather(city: str) -> str:
    """Get current weather for a city"""
    url = f"https://wttr.in/{city}?format=%C+%t+%w"
    with urllib.request.urlopen(url, timeout=10) as r:
        return r.read().decode("utf-8")

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

This shows the full loop: user asks → AI picks the tool → tool fetches data → AI answers. Follow the same pattern to add your own APIs.

6. stdio vs SSE

7. One Server, Many Tools

A single server can register unlimited tools. Clients auto-discover them and pick the right one per request. Good tool descriptions make the AI use them correctly.

FAQ

Q: ModuleNotFoundError: mcp? A: Run pip install mcp, and make sure the script runs with the same Python environment. Q: Nothing shows in the terminal? A: Normal for stdio servers; output appears when a client calls a tool. Q: MCP vs a normal API? A: With MCP the AI client discovers and calls tools itself—no per-tool integration code, and it works across clients. Q: Can I deploy remotely? A: Yes, using streamable HTTP transport with your own auth and firewall; see the official docs.

Note: SDK APIs may evolve; check modelcontextprotocol.io if something fails.

❓ FAQ

ModuleNotFoundError: mcp?

Run `pip install mcp`, and make sure the script runs with the same Python environment.

Nothing shows in the terminal?

Normal for stdio servers; output appears when a client calls a tool.

MCP vs a normal API?

With MCP the AI client discovers and calls tools itself—no per-tool integration code, and it works across clients.

Can I deploy remotely?

Yes, using streamable HTTP transport with your own auth and firewall; see the official docs. > Note: SDK APIs may evolve; check modelcontextprotocol.io if something fails.