Before going live on RouterBrain.

Configuration

  • Prod URL and keys only on servers/secrets (Security best practices)
  • baseURL includes /v1 (OpenAI surface)
  • Anthropic uses https://ai.win10030.com/anthropic/v1
  • Production and non-production keys and Base URLs are not mixed

Network

  • HTTPS only
  • Egress IPs on allow list (if used)
  • Proxy SSE buffering disabled (Streaming)
  • Read timeout ≥ max expected generation

Reliability

  • Backoff on 429/5xx (Errors)
  • Streaming handles disconnect and [DONE]
  • Timeouts / AbortSignal on critical paths
  • Idempotent tool side effects

Observability

  • x-trace-id on every call
  • Redacted logs (no full keys / raw prompts)
  • Metrics on 401/403/429/5xx and P99 latency
  • Support template ready (Support information)

Functional checks

  • GET /models works
  • Non-stream and stream chat pass
  • model ids exist and are entitled
  • Files/PDF path if used
  • Embeddings via HTTP if used

Capacity

  • Concurrency limits and queues
  • Load-tested large contexts/attachments

Related

Architecture · Key lifecycle