Podcast Search Engine and MCP Server

I built a search engine over a library of more than a thousand product and technology podcast episodes. Transcripts were chunked into tens of thousands of passages, embedded, and indexed in Pinecone, so a question returns the passage that actually answers it rather than the episode whose title happens to match the keywords.

The more interesting half was turning that product into something another AI could use. Using FastMCP I exposed the same engine over the Model Context Protocol, which meant deciding what a capable client should be handed rather than what a human clicks.

That split into four kinds of surface. Tools let the model act: list the library, search across the corpus, pull a full transcript to summarise, and save or recall notes so context survives between sessions. Resources expose the catalogue as attachable context. A prompt packages a repeatable research brief. Finally the server sits behind bearer token authentication, since an MCP endpoint is a live API and shipping one without auth is shipping an open database.

Connected to an assistant, it stops being a website I search and becomes a library the model reaches into mid-conversation: pulling together what the corpus says about a topic like pricing, then keeping a note on it for next time.