7 Things I Learned Building a Rate-Limited MCP Server in Elixir (mikenotthepope.com)

The author built "Hit Me DB One More Time," an MCP (Model Control Protocol) server in Elixir to demystify how AI assistants call external tools. The post condenses seven practical lessons: an MCP server is just a regular program that advertises tools at startup, clients (e.g., ChatGPT, Claude Code) already include MCP support, communication uses JSON-RPC over stdin/stdout or HTTP, and the LLM decides which tool to call while the host actually executes it. The request flow is straightforward—announce tools, LLM requests a call, host issues JSON-RPC to the server, server returns results, and the LLM responds to the user. Technically, the Elixir server demonstrates a plugin/middleware pipeline that can observe, modify, short-circuit, or forward requests—critical for production safety and efficiency. The example uses a SQLite DB with 100 sample records and a paginated get_records tool, plus middleware implementing rate limiting (5 requests per 10 seconds), caching (30s TTL), and logging. The takeaway for AI/ML practitioners: MCP makes integrating assistants with internal data and actions predictable and extensible, but servers must defend resources with middleware (rate limits, caches, auth) since LLMs will call tools repeatedly if allowed.