llmgw
Docs/Limitations

Product contract

What llm-gateway does not claim.

Explicit limits make exact routing and compatibility more useful than broad promises.

This page is part of the product contract, not a future-feature checklist.

Scale

  • Single-node SQLite control plane.
  • No clustered state, distributed locks, horizontal scaling, or HA scheduler.
  • No hyperscale throughput target.

Protocol scope

  • Core OpenAI surfaces, not every API or field.
  • Core Anthropic Messages/counting, not every beta feature.
  • Unsupported or lossy translated fields fail closed.
  • No versioned management OpenAPI yet.

Routing

  • Ordered failover, not weighted, round-robin, least-used, cost-first, or quota-aware scheduling.
  • Streaming moves only before the first byte.
  • Embeddings, audio, image, and video select one eligible target.
  • Provider circuit state is process-local.

Providers

  • Registry support does not guarantee account/model/deployment/region access.
  • Google chat sends message text and function tools only; images do not reach the model.
  • Native image/video generation is currently Google-capable-provider specific.
  • No production quota adapter drives routing; unknown quota remains unknown.
  • Estimated cost is not invoice reconciliation.

Client confidence

Claude Code, Codex, and Copilot CLI have fixture-backed profiles and a real-UAT procedure, not blanket certification of every future release.

Identity and credentials

  • Generic system API-key fallback is not project-binding-gated.
  • Codex and personal Copilot OAuth are human-private.
  • No Claude personal OAuth, cookie extraction, MITM, or stealth sessions.
  • secrets.json can retain plaintext compatibility credentials.

Console

  • Shared Copilot binding APIs lack a complete primary-console editor.
  • Not every provider form exposes force_api_support.
  • Portal capability-specific media actions are not all separate user routes.

Operations and security

  • Backups are operator-triggered and local.
  • Request logs are excluded from backup.
  • Checksums do not authenticate archives.
  • Health is liveness, not provider readiness.
  • Alert delivery requires an external worker.
  • Included Caddy/Compose is not turnkey SSO.
  • SSRF blocking is intentionally absent.

Removed from scope

No universal translation IR, 100-provider target, quota pools/borrowing, trace-console expansion, streaming account selection, or public hosted gateway service.