FastMCP

FastMCP is the framework under every MCP server I build, and the skill is knowing which decisions the framework leaves to you. The framework handles the protocol; the server's worth is decided in what its tools say and refuse. Three servers exercise that: a durable agent-memory server published to PyPI and listed in the official MCP Registry, whose async tool layer fronts thread-scoped storage with full-text, semantic, and hybrid retrieval; a production system-of-record server that turns a project-tracking platform into agent-callable tools, plus a second, deliberately narrow endpoint exposing only the search and fetch pair ChatGPT's Deep Research connector accepts; and an HR-platform server in development, designed fail-closed around a two-tier sensitivity model before the first tool shipped.

The recurring craft across them: tool surfaces curated for an agent's context budget rather than generated one-to-one from an upstream API — responses trimmed to what the caller needs with raw payloads one parameter away, errors written as corrections an agent can act on rather than status codes, per-user credentials whose responses are never cached, and partial degradation that marks the inaccessible corner instead of failing the whole call. Servers run across stdio and streamable HTTP transports, into Claude Code, agent runtimes, and gateway deployments alike. The three servers are in the Related Projects panel above.