Models

Your model, your key, your endpoint.

Model choice is infrastructure, not architecture. Changing it should be configuration, and your key should never travel further than it must.

Why model lock-in bites

Use your provider account, with per-run accounting of what each run cost.

Paste an OpenAI-compatible URL; the save-time check actually calls it, so only an endpoint that answered is stored.

Model traffic never crosses into the browser host — a deliberate boundary, not an implementation detail.

Whatever model you choose, results are validated against your schema before you see them.

A goal and a schema over HTTP — no selectors to maintain and no browser fleet to run.

“Add a model in the dashboard with your own key or endpoint; runs record which model resolved for each one.”

Every run stamps the model actually used, for audit.

Are my API keys stored safely?

Keys are held in a secrets manager and referenced, never stored in the database. They also never cross into the browser environment.

Can I use a self-hosted model?

If it speaks the OpenAI-compatible API over HTTPS and is reachable, yes — and it is validated at save time by a real call.

Which providers are supported?

A provider is enabled only once it passes a live contract test with real keys, so the list reflects what actually works rather than what we intend.

Start with one task. Watch it run, take the wheel whenever you want, and keep the ones that earn their place.