Optional. When Puckie doesn't recognise a request, a small, fast model decides whether you asked a question or gave a command it can't do yet, so you hear “I can't do that yet” at once instead of waiting for an answer.
| Choice | What it means |
|---|---|
| Off | The answering model handles everything Puckie doesn't recognise. Fine for most people. |
| Laya · this Mac | Private and very fast (a few milliseconds). Runs with Puck's built-in model. Nothing to set up. |
| Jev · online | TypeSafe's model, fast and accurate, but what you say is sent online. Needs a key. |
| Custom | Your own server that works the way Jev does. |
Puckie only refuses when the model is almost certain. A question it's unsure about still gets an answer.
Jev is sold directly and through gateways you may already use. Pick one, get a key, and choose it under Through.
| Through | Where to get the key |
|---|---|
| TypeSafe | console.typesafe.ai |
| Vercel AI Gateway | vercel.com → AI Gateway → API Keys |
| OpenRouter | openrouter.ai/keys |
| Cloudflare | dash.cloudflare.com: your Account ID, and an API token allowed to use Workers AI |