"Just upload your data for better AI" is a common pitch. For a personal starred library, it is the wrong default.

Starcat's AI features are built around BYOK — bring your own keys — so models run through your provider settings instead of quietly shipping your notes into someone else's training-shaped void.

What BYOK actually buys you

  1. Choice — pick the provider and model that fit cost, latency, and quality
  2. Boundary — keys live in Keychain-backed storage on your Mac; the app does not invent a second identity for your library
  3. Accountability — usage and failures are yours to see, not buried in a opaque pooled meter

Pro unlocks the capability. BYOK keeps the control surface honest.

Local library, remote models

The split is intentional:

  • Local: Stars cache, tags, notes, status, search indexes, RAG chunks
  • Remote (optional): inference and embeddings when you ask for them

That means AI is a tool you switch on for a job, not a cloud that owns the filing cabinet.

Practical advice

  • Start with one provider you already trust for other work
  • Prefer models you can afford to run on "suggest tags for this repo" without flinching
  • Keep the confirmation rule: suggestions are drafts until you accept them

If a product only works when it holds your API relationship and your corpus, you are renting both. Starcat tries to rent neither by default.