Whether the relay is up, and whether each AI provider behind it is answering right now.
AI backends
The relay asks each provider directly, about once a minute, on an endpoint that provider does not charge for. This is a real answer from the provider, not a guess from traffic: a backend reads alive only if it responded.
What the states mean
Alive — the provider answered the last check. A provider that is answering but throttling us (429) is still alive; it is marked rate limited, and calls to it may be slow or refused for a while.
Down — it did not answer, answered with an error of its own, or rejected the relay's key. No call can be served on it until that changes.
Not configured — the relay holds no key for this provider, so it is never routed to. Nothing is broken; it is simply not one of the backends today.
Not checked yet — the relay restarted less than a minute ago and has not asked yet. It is not a claim that anything is wrong, and not a claim that anything is fine.
Rate limits and billing rejections seen on real forwarded calls are shown too, from the relay's rolling window of recent calls. Those say what customers actually hit, which a probe on a free listing endpoint cannot see.