You want NotebookLM's two-AI-hosts-discussing-your-content trick, but on your own infrastructure — your API keys, your storage, no Google account in the loop. That's exactly the shape of this project.
Unlike self-hosting an LLM, this pipeline is glue, not inference — the heavy lifting is API calls. It runs happily on GitHub Actions' free tier or any $0 leftover machine; there's no GPU, database, or always-on server. npm run setup provisions storage, deploys the player, and prints your private podcast feed URL.
No interactive mode (you can't interrupt the hosts to ask questions), and no multi-document research grounding. What you get instead is automation: a daily audio digest of your sources and full audiobooks of your epubs, delivered to a normal podcast app.
No. It's an orchestration pipeline: TTS and LLM are API calls. You can point it at a local model if you run one, but it's optional.
In your own object-storage bucket with a private feed URL. Delete the bucket and every trace is gone — it's your data in the most literal sense.
Storage and compute fit in free tiers; LLM rewriting costs a few cents a day and is optional. Effectively $0.