inference-providers

Model registries like models.dev exist, but inference-providers tracks provenance: lab-owned model facts live separately from provider-owned serving details (auth, endpoints, protocols), stored as two layers of TOML that pass validation gates before being emitted as typed artifacts. Structured reasoning/thinking parameter specs mean tools don't guess which knobs each provider supports.

Consumption happens through an SDK — loadCatalog, resolveModel, authHeaders, buildReasoningParam — backed by a generated docs site deployed via GitHub Pages.