Provider access stays separate from application control.
01
Preserve the Messages boundary
Claude clients keep the Anthropic Messages request shape while the gateway applies the approved workload, route, model, and provider account. Streaming uses the same governed path and must be validated for the selected model before rollout.
02
Separate identity from provider access
Each application receives a scoped AI Gateway HQ workload key. The Anthropic credential remains encrypted and write-only, so rotating a provider key does not require placing it on every developer machine or service.
03
Fail over only when the request is compatible
Health, quota, policy, budget, and capability checks run before a target is eligible. A secondary target is not treated as safe merely because it can generate text; it must support the request's exact protocol boundary.
Current shared controls
Status determines what can be configured.
Write-only, encrypted BYOK for implemented provider connections Stable model aliases and protocol-eligible route targets Organization and key budgets, RPM, TPM, and concurrency hard limits Observe, Shadow, Enforce, reason codes, and metadata-only evidence
Beta connections require customer validation against the exact model, payload, streaming mode, region, and provider account before production use.
Applications keep a stable gateway URL and model alias while an administrator changes eligible provider accounts and models.
Implemented boundary
Only documented protocols become eligible.
Choose Anthropic, enter the API key once in the write-only credential form, verify account access without generating tokens, and add only Messages-compatible targets to the route.
POST /v1/messages Anthropic-compatible streaming GET /v1/models connection test
Compatibility is bounded to the provider's current published interface. Review the provider documentation before approving a production model.