Patricio Martins Logo
Back to articles

Best Practices for Building MCP Servers with a Focus on FastMCP

1. Introduction to the Model Context Protocol (MCP) and FastMCP

Connecting Large Language Models (LLMs) to external systems is a growing challenge. The Model Context Protocol (MCP) standardizes this communication, letting LLMs access context and use tools to act in the real world. FastMCP is a robust framework that makes building MCP servers in Python faster and safer.

1.1. The role of MCP in LLM interactions

MCP defines how applications provide context and expose tools to LLMs. It enables complex AI workflows by connecting models to diverse data sources and actions through a consistent API.

MCP follows a client–server architecture with three pieces:

  • MCP Hosts: Apps that consume resources and tools via MCP (e.g., IDEs, AI apps)
  • MCP Clients: Maintain connections to servers
  • MCP Servers: Lightweight programs that expose resources and capabilities

1.2. FastMCP: A Pythonic approach

FastMCP abstracts protocol complexity and server management so developers can focus on tools and value. Version 2.0 provides the “fastest path to production,” including proxying, server composition, OpenAPI‑based generation, and rich client‑side features.

1.3. Getting started

uv add fastmcp
# or
uv pip install fastmcp
# or
pip install fastmcp
fastmcp version

Minimal server example:

from fastmcp import FastMCP, mcp

mcp_server = FastMCP(
  name="MyFirstMCPServer",
  instructions="This server provides basic math tools and a greeting resource."
)

@mcp_server.tool()
def add(a: float, b: float) -> float:
    return a + b

@mcp_server.resource("greeting://general")
def greeting() -> str:
    return "Hello! How can I help?"

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

2. Essential components of a FastMCP server

Tools are the primary interface for actions; resources expose read‑only data; prompts structure reusable interactions. Use clear names and concise descriptions.

3. Performance optimization

The biggest performance killer is token overhead from verbose tool schemas. Optimize by:

  • Using concise descriptions
  • Removing redundant examples
  • Linking to external docs when possible
  • Loading tools dynamically where appropriate
  • Bundling related tools

Tune infrastructure as usual: strong CPUs/GPUs, NVMe SSDs, containerization, and observability.

4. Security best practices

  • Prefer OAuth 2.1 + OIDC; avoid long‑lived static tokens
  • Enforce RBAC at the tool level
  • Rotate secrets regularly; use vaults (e.g., Vault, AWS Secrets Manager)
  • Add DLP and PII detection for traffic
  • Validate inputs rigorously; sandbox local command execution
  • Use end‑to‑end encryption; consider mTLS for sensitive backends

5. Design and maintenance

  • Design from user workflows downward, not from endpoints upward
  • Keep tool naming/documentation clear; logs centralized with alerts
  • Plan for resilience: backups, DR drills, vulnerability management
  • Treat configurations as code; validate YAML/JSON in CI

This guide summarizes practical lessons to build safe, fast, and maintainable MCP servers with FastMCP in production.