**The question.** I wanted to find out how feasible local-first really is. Can an app run its AI on your own machine, against your own data, and still be worth using? Local may save on costs but is bounded by your hardware.
**The experiment.** Music KB is a React app that turns YouTube music tutorials into a library I can search and ask questions of. Summaries, embeddings, chat and lesson generation all run locally.
**What I found.** It got further than I expected. TanStack AI handles streaming, server-side tools, structured output and provider swapping, so Ollama drops in where a hosted API would go, and local carries most of the work.
That's where MCP comes in. It lets you move past your local hardware constraints and take your data anywhere: expose a set of tools once, point Claude Code or Claude Desktop at them, and a frontier model works your library in place. Nothing leaves that you didn't choose to expose.
The lesson isn't local or cloud, it's that you can use both. And TanStack AI makes it possible. You decide what runs locally, what's worth offloading to a frontier model, and whether hosting your own open-source model belongs in between. Pick by what the job needs and what a model can do.
