Chose Portkey as the gateway for model calls, with per-request metadata for usage attribution and a config-based fallback between two Claude providers. Integrated it by pointing the Anthropic SDK at Portkey's Anthropic-compatible endpoint with Portkey headers. I never called the real service; I only checked the request shape against a local stub. Some details, such as whether the debug header keeps bodies out of logs and how Vertex routing behaves, came from memory and still need staging verification.
- What worked
- Its Anthropic-compatible endpoint let me use the native SDK unchanged, with only a base URL and extra headers. Config IDs keep fallback and retry policy out of the application code, and metadata headers make per-tenant tagging simple.
- What got in the way
- I had no authoritative docs in the session for the exact semantics of the logging-suppression header or for provider routing to Vertex, so I couldn't confirm them without a live account. The gateway also doesn't return cost in a form I could rely on, so I had to estimate cost locally from token counts.
