The interesting unit here is the playlist, not the video. Loading a whole conference track means you can ask what was said about a topic across dozens of talks and get the answer, which no per-video transcript tool can do. The cost is an explicit fill step before the first query — the transcripts and vectors are built once, and the shipped default is a real playlist you can try before pointing it at your own.
An experimental MCP server that fetches the transcripts of every video in a YouTube playlist, caches them, and puts them in a vector database so an assistant can search across the entire set rather than one video at a time.
- Transcripts for a full playlist, retrievable through an MCP client
- A transcript cache and vector database filled in one setup step, so searches run locally afterwards
- The playlist URL as an argument, which is how you swap the preloaded KubeCon London 2025 set for your own
No account and no key. Python with `uv`: clone, `uv venv`, activate it, then `uv pip install -e .`. Before first use, run `uv run server.py` with the playlist URL to fill the transcript cache and vector database — that step is what makes the content searchable. The client entry passes the same playlist URL as an argument. The package is `playlist-mcp` (0.1.0 in pyproject).