The model card is the substance: pricing per call and per million tokens, context window, maximum output, modalities and both p50 and p95 latency in one response, which turns a choice between two models into a comparison you can defend. Routing is addressable at three levels — the workspace auto-router, a named router, or a provider and model pinned directly — so a call can be fixed where it matters and left automatic where it does not. The provider list exists because the filters take provider IDs, and those are not reliably guessable from display names.
A client for OrcaRouter's model catalogue with a single-turn chat call across it.
- A single-turn chat request with a system prompt, temperature and token ceiling, defaulting to the workspace auto-router and addressable either as a named router or as a provider and model directly.
- The model catalogue with id, name, description, owner, context length, supported endpoint types and pricing both per token and per million tokens, filtered by provider, capability or minimum context.
- A model card carrying the long description, per-call and per-million pricing, context window, maximum output, input and output modalities, supported endpoints, p50 and p95 latency, and release information.
- The provider list with provider IDs, display names and model counts, which is where the valid filter values come from.
An OrcaRouter workspace to route the calls through.
One command plus a key — npx -y @orcarouter/mcp, then supply credentials