Endpoints
OpenGateway exposes OpenAI/Anthropic shapes where they exist and native
passthrough where they don’t, on every lane (/oss, /hf, /frontier, /pro).
Prefix any path below with a lane, e.g. /oss/v1/chat/completions.
| Task | Path | Wire shape | Streaming |
| --- | --- | --- | --- |
| Chat | /v1/chat/completions | OpenAI Chat | SSE |
| Messages | /v1/messages (+ count_tokens) | Anthropic Messages | SSE |
| Responses | /v1/responses | OpenAI Responses | SSE |
| Completions | /v1/completions | OpenAI legacy | SSE |
| Embeddings | /v1/embeddings | OpenAI Embeddings | — |
| Rerank | /v1/rerank | Cohere/Jina-style | — |
| Images | /v1/images/generations | OpenAI Images | — |
| Video | /v1/videos/generations | native (url/b64_json) | — |
| Audio | /v1/audio/speech, /v1/audio/transcriptions | OpenAI Audio | — |
| Models | /v1/models | dual-shape | — |
Two HF routing families
Section titled “Two HF routing families”Under the hood, the Hugging Face OpenAI-compatible /v1 surface is chat-class
only (chat completions, Responses, and /v1/models). Everything else
(embeddings, images, video, audio, rerank) is served via per-provider routed
paths that OpenGateway normalizes into the OpenAI/Anthropic shapes you see
here. You don’t have to care which family a request uses — the lane and endpoint
are all you specify.
For the full interactive schema, see the API reference.
Ask about configuring OpenGateway — lanes, base URLs, client setup, model choice, or an error you hit. Answers are grounded in the docs.