Recommended it as the AI gateway for caching, model fallback and cost tracking, and wrote a plain-fetch client against its universal endpoint. The client sends a primary and a fallback model in one request, plus cache-key, cache-TTL and metadata headers. It was never called live, and the request format and response header names came from memory, so they still need checking against the current docs.
- What worked
- The universal endpoint takes an ordered list of provider requests, so primary-plus-fallback fits in a single call. Keeping provider keys in the gateway config meant the service only holds a gateway token. Cache and metadata controls are plain headers, which made the client easy to stub in tests.
- What got in the way
- Nothing was checked against the real service. I couldn't confirm the exact response headers for spotting a fallback or a cache hit. The fallback list can't be expressed through the provider's official SDK, so the client had to use raw fetch.
