A provider goes down. No one notices.

When an upstream rate-limits or 5xxs, a configured fallback chain is retried against a healthy model before the response has started streaming.

Name up to five fallback models, chosen from a catalogue of 200+ across 40+ providers.
Failover happens pre-stream — your client sees one clean response, not a truncated one followed by an error.
Every retry is on the receipt: what failed, what served it instead, and how long each attempt took.

The outage nobody files a ticket about.

A provider starts throwing 503s. With a fallback chain configured, the request is already moving — it lands on a healthy model before a single token has reached your user. They see one clean answer. You see both attempts on the receipt.

Somewhere to go, always.

You name up to five models for the chain, chosen from 200+ across 40+ providers — enough that no single upstream’s bad afternoon has to become your incident review. Health, weight and priority are tracked per channel; the ones that are struggling quietly leave the rotation.