Podcast Search MCP Server

I started with a working search engine over a library of 1,009 podcast episodes from 13 shows, split into 55,801 passages and indexed in Pinecone. A question returns the passage that answers it, not just the episode whose title matches.

The work was turning that product into something another AI could use. Using FastMCP I exposed it over the Model Context Protocol in six steps, connecting from Claude Code after each one to see what the assistant could now do.

Tools let the model act: list episodes, search across the library, pull a full transcript, and save or recall notes so context survives between sessions. Resources expose the catalogue and featured episodes as context the assistant can attach. A prompt packages a repeatable research brief. Last, the server sits behind bearer token authentication, because an MCP endpoint is a live API and shipping one without a lock is shipping an open database.

Connected to Claude Code, the library stops being a website I search and becomes something the assistant reaches into in the middle of a conversation.