AI For Dev

summary-cole-medin-mcp-server-template

2 min read

The ULTIMATE Guide to Building Your Own MCP Servers

Source: YouTube — Cole Medin, published 2025-04-13 Tools: MCP (Python SDK / FastMCP), Mem0, Docker

Summary

Cole Medin’s comprehensive guide to building MCP servers in Python using FastMCP. Walks through the SDK fundamentals (@mcp.tool decorator, lifespan pattern for singleton clients, transport options), then builds out a full Mem0 MCP server with three tools (save_memory, get_all_memories, search_memories) deployed via Docker. The template is published at github.com/coleam00/mcp-mem0 and is the wiki’s recommended starting point for anyone building their first MCP server.

Key facts

  • SDK install: pip install mcp; FastMCP is the high-level API of choice
  • Transports: SSE (HTTP, remote, port 8070 default) vs stdio (local subprocess, faster). Choose stdio for IDE-embedded use, SSE for shared servers.
  • Lifespan pattern: singleton database/API client created at server start, passed via context to all tool calls — avoids reconnect-per-call overhead
  • Tool decorator: @mcp.tool + docstring. The docstring is what the LLM sees when picking tools — write it like a tool description, not a Python docstring
  • Mem0 template: 3 tools, deployable via Docker or python server.py, repo github.com/coleam00/mcp-mem0

Why it matters

This is the wiki’s canonical “how to build an MCP server” reference — until now, MCP coverage was about consuming MCP servers (Stagehand, Task Master, Brave Search, Perplexity). Cole’s video is the wiki’s first source on the production side. Pairs with the Context7 walkthrough (also Cole, also this batch) for the “what does a great MCP server look like” companion piece.

The lifespan + stdio combination is the part most beginners miss — they default to SSE because the docs lead with it, then wonder why every tool call reconnects. Cole’s framing fixes that for any future builder reading the page.

See Also

Graph View

Backlinks