Wyolet Relay puts a single OpenAI- and Anthropic-compatible endpoint in front of every provider you use. Pool your own API keys for automatic failover and higher effective rate limits, see exactly what every request costs, and run the whole thing on your own infrastructure — a drop-in for the SDK code you already have.

Quickstart

Start a full relay — API, admin UI, database, and a pre-seeded model catalog — in one command:

docker run -p 8080:8080 -p 8081:8081 wyolet/relay:standalone

Open the admin UI at http://localhost:8081, then let the setup wizard walk you through adding a provider key and minting a relay key. Now call it like the OpenAI API:

curl http://localhost:8080/openai/v1/chat/completions \
  -H "Authorization: Bearer <your-relay-key>" \
  -H "Content-Type: application/json" \
  -d '{"model":"gpt-4o","messages":[{"role":"user","content":"hello"}]}'

That's it. The full walkthrough, configuration, and production deployment guides live at docs.wyolet.com.

Features

• One API, every provider. OpenAI- and Anthropic-shape endpoints in front of OpenAI, Anthropic, Bedrock, Vertex, Azure, Ollama, Groq — anything speaking either wire format. No code changes to switch upstreams.
• Disposable, rate-limited keys. Mint relay keys scoped to whatever limits you set. Hand them out freely — even if one leaks, the damage is capped at those limits and your real provider keys are never exposed.
• Pool accounts and providers. Combine many keys, accounts, or providers into one pool behind a single endpoint. Relay load-balances and fails over across them, so per-account rate limits stop being your ceiling.
• Per-key access control. Decide exactly which models and providers each relay key may reach — allow or deny at the key level via policies.
• 400+ models, open catalog. Ships knowing 400+ models out of the box, and the catalog is open and extensible — we add hosts and models on demand.
• Batch processing (in progress). Batch requests against any provider — Relay simulates batching where there's no native API, and routes through the native one (OpenAI, Gemini, Anthropic) where it exists, passing the cost discount straight through. Configure a webhook to fire when a batch completes.
• Proxy mode. Point Relay at a provider with your own upstream key and use it as a transparent proxy — no policy enforcement, just full usage, cost, and payload logging.
• Usage & cost tracking. Every request is metered and stored in Postgres or ClickHouse. Optional full request/response payload capture (off by default).
• Metrics & logs. First-class Prometheus /metrics and structured JSON logs. (OpenTelemetry tracing is on the way.)
• Self-hostable, built for scale. Bring your own keys; nothing phones home. Sub-2 ms added latency, thousands of requests/sec per pod, Kubernetes-native.

How it works

Relay runs two listeners: a data plane that accepts your inference requests and a control plane that serves the admin UI and API. Each request is authenticated by a relay key, matched to a policy that decides which models and providers it may reach, rate-limited, and routed to a healthy upstream key from the pool — then streamed straight back to you. Provider, model, and pricing data comes from an open, versioned catalog, so a fresh container already knows hundreds of models.

Want the full architecture, API reference, and configuration? → docs.wyolet.com

Commercial support

Relay is Apache-2.0 — free to use, self-host, and build on, in commercial and closed-source products alike. Want managed hosting, enterprise builds, or priority support instead of running it yourself? We're happy to talk: business@wyolet.com.

Contributing

Issues and pull requests are welcome. See CONTRIBUTING.md for the build, test, and PR workflow.

License

Apache-2.0. Use it in anything — commercial, closed-source, hosted, or embedded — no copyleft strings attached. See Commercial support if you'd rather we run or support it for you.