If you've been watching the AI developer tooling space, you've noticed Model Context Protocol (MCP) go from a niche Anthropic proposal to the de-facto standard for connecting LLMs to external data and tools. By early 2026, every major AI provider — OpenAI, Anthropic, Microsoft, AWS — had adopted it, and there are over 10,000 public MCP servers available.

This guide skips the theory. You'll build a working MCP server in Python, test it locally, and deploy it to production. Total time: about 45 minutes.

What is an MCP Server? (30-second recap)

An MCP server exposes three primitive types to any MCP-compatible AI client (Claude Desktop, Cursor, VS Code Copilot, etc.):

Tools — functions the LLM can call (e.g., query a database, send an email)
Resources — read-only data the LLM can fetch (e.g., a file, a live dashboard)
Prompts — reusable prompt templates with parameters

Think of it as "a USB-C port for AI" — one protocol, any client.

Prerequisites

Python 3.11+
pip or uv (recommended)
Claude Desktop installed (free) — or any MCP client
Basic Python familiarity

Step 1 — Project setup

Create a new project and install the MCP Python SDK:

mkdir my-mcp-server && cd my-mcp-server

python -m venv .venv && source .venv/bin/activate

pip install mcp httpx

The mcp package gives you FastMCP, a decorator-based server that handles all protocol boilerplate.

Step 2 — Define your first MCP tool

A tool is a Python async function decorated with @mcp.tool(). The docstring is the tool description — the LLM reads it to decide when to call this tool. Write it clearly.

Key things to notice:

The docstring is the tool description — the LLM reads it to decide when to call this tool
Type hints become the parameter schema automatically
The function must be async

Step 3 — Add a resource

Resources are read-only data sources fetched by URI, not called with arguments. A common use case: expose a company knowledge base or a live config file.

Use @mcp.resource("config://app/settings") to expose configuration data. The URI is how MCP clients reference this resource.

Step 4 — Add a reusable prompt

Prompts let you package prompt templates that users or the LLM can invoke by name. Use @mcp.prompt() with a GetPromptResult return type.

Step 5 — Test locally with Claude Desktop

Add the server entry to Claude Desktop's config file at ~/Library/Application Support/Claude/claude_desktop_config.json. Restart Claude Desktop. You should see the hammer icon in the chat input — that means Claude has detected your MCP tools.

Step 6 — Deploy to production (Streamable HTTP)

For local use, stdio transport works fine. For production, switch to Streamable HTTP:

python server.py http

Your server is now reachable at http://localhost:8000/mcp. Deploy behind an HTTPS reverse proxy for production. Use Docker for containerized deployment.

Common mistakes to avoid

Writing vague tool descriptions — the LLM uses your docstring to decide when to call the tool
Synchronous blocking calls — use httpx.AsyncClient, not requests
Returning unstructured text — return structured JSON so the LLM can reliably parse results
One giant server for everything — keep servers focused and composable

MCP vs REST API

Use MCP when you're building something an AI agent will consume. Use REST when a human application is the primary consumer. Many teams build both — a REST API for the product and an MCP server that wraps it for AI agents.

FAQ

Do I need to use Claude to build an MCP server? No. MCP is an open standard. Any MCP-compatible client works — Claude Desktop, Cursor, VS Code GitHub Copilot, Zed, and many more.

Is MCP production-ready in 2026? Yes. Streamable HTTP solved the stateful session and horizontal scaling problems. AWS, Azure, and GCP all support MCP-native deployments.

Can I use MCP with CrewAI or LangChain? Yes. LangChain has langchain-mcp-adapters. CrewAI supports MCP tool integration via its tool registry.

What's the difference between an MCP tool and a resource? Tools are callable functions with side effects. Resources are read-only data sources identified by URI.

How do I secure a remote MCP server? Use HTTPS, Bearer token auth, rate-limiting, input validation, and sandboxed containers.

How to Build Your First MCP Server in 2026 (Step-by-Step)