One router.
Every key you already hold.
Sabi dispatches to a provider over an OpenAI-compatible chat interface. Each page below states the protocol, the base URL, and the environment variable that resolves your key — the same three facts the gateway uses at dispatch.
8 adapters
OpenRouter
An aggregator: one OpenAI-compatible key reaches many vendors, including a set the catalogue reports as zero-priced.
Hugging Face
The Inference API is OpenAI-compatible at the router host, so the same chat body and Bearer token work as for every other provider here.
Groq
Fast inference on an OpenAI-compatible surface. Where the catalogue reports a zero price, it is a free-lane candidate.
NVIDIA NIM
NVIDIA NIM serves open models over an OpenAI-compatible endpoint.
Google / Gemini
Native Gemini. Model ids are provider-scoped and differ from the ids the same models carry on an aggregator, so one catalogue entry does not address both.
Mistral
Mistral’s own inference API, OpenAI-compatible at its base URL.
OpenAI
The OpenAI API, OpenAI-compatible. Most entries are priced, so they route only when paid inference is enabled.
Ollama
A runtime you run yourself. Sabi dispatches to it over the same OpenAI-compatible chat surface, pointed at your own machine.