yym68686/uni-api

uni‑api – A unified, self‑hosted gateway for many LLM providers

What it is

• An open‑source server that exposes the standard OpenAI HTTP endpoints (e.g., /v1/chat/completions, /v1/images/generations, /v1/embeddings, etc.) but forwards the requests to a configurable list of backend LLM services.
• It lets you run a single API surface while mixing and matching providers such as OpenAI, Anthropic, Gemini, Vertex AI, Azure, AWS Bedrock, Cohere, Groq, Cloudflare Workers AI, OpenRouter, and others.
• All configuration is done via a plain api.yaml file (or a URL supplied in CONFIG_URL), so no separate UI is required.

Key capabilities

• Provider‑agnostic routing – map any OpenAI‑style request to the appropriate backend URL and API key.
• Load‑balancing – three modes:
  1. Weighted per‑channel balancing.
  2. Round‑robin (sequential) per‑channel balancing.
  3. Automatic round‑robin across multiple API keys within a channel.
• Automatic retry & cooling – if a backend returns an error, the request is retried on the next channel; failing channels are temporarily “cooled” and excluded until the cooldown expires.
• Fine‑grained timeouts – set different request timeouts per model.
• Permission & rate‑limit controls – per‑API‑key model whitelists, wildcard patterns, and configurable request caps (e.g., 2/min, 5/hour).
• Tool and image support – native tool‑calling and image‑generation endpoints are passed through for providers that support them (Anthropic, Gemini, Vertex, etc.).
• Moderation – optional OpenAI‑style content moderation that can block disallowed user messages before they reach the backend.
• Custom request tweaks – per‑model overrides for request body fields, header injection, proxy settings, and maximum payload size.

How to run it

1. Prepare api.yaml – list your providers, their base URLs, API keys, and the models you want to expose. The README includes a minimal example and a very detailed advanced example.
2. Deploy – either:
   • Mount the file into the Docker container as api.yaml, or
   • Set CONFIG_URL to point at a remote YAML file; the container will download it on start.
3. Start the container – the image is published on Docker Hub (yym68686/uni-api). A one‑click “Deploy to Fugue” button is provided for quick cloud deployment.
4. Consume the API – call the usual OpenAI endpoints on the host (e.g., http://localhost:8000/v1/chat/completions) using any of the API keys you defined in the api_keys section.

Who it’s for

• Individuals or small teams who want a personal “proxy” that aggregates many paid LLM services without building a custom UI.
• Developers needing a consistent OpenAI‑compatible interface while experimenting with different providers or fallback strategies.
• Anyone looking for simple YAML‑based configuration of load‑balancing, rate‑limiting, and model‑specific tweaks.

Links

• Telegram discussion group: https://t.me/uni_api
• Docker Hub page (pull counts, tags): https://hub.docker.com/repository/docker/yym68686/uni-api
• One‑click deployment: https://fugue.pro/new/repository?repository-url=https%3A%2F%2Fgithub.com%2Fyym68686%2Funi-api

---

All details above are taken directly from the repository’s README; no additional features have been inferred.