One OpenAI-shaped front door for a zoo of LLM backends

What it does

One-api is an API gateway and key-redistribution layer that sits between your application and a long list of LLM providers—OpenAI, Azure, Anthropic, Google, DeepSeek, and a parade of Chinese models including Baidu, Alibaba, ByteDance, and Tencent. It exposes a standard OpenAI-compatible API to your clients, then routes, load-balances, and meters requests across whichever backends you configure. The project ships as a single binary with a web dashboard for channel and token management.

The interesting bit

The real workhorse is the channel system: you can register multiple accounts or endpoints for the same provider, group them, set per-model rate multipliers, and let the system fail-over or round-robin between them. It also turns the whole affair into a small business if you need it to—supporting redemption codes, user invitation rewards, quota billing in USD, and fine-grained token controls like IP allowlists and model restrictions.

Key highlights

• Supports a wide mix of Western and Chinese providers—OpenAI, Azure, Claude, Gemini, Mistral, Groq, Ollama, DeepSeek, Baidu, Alibaba, iFlytek, Zhipu, 360, Tencent, Moonshot, and others.
• Presents a unified OpenAI-compatible API to downstream clients regardless of the backend vendor.
• Built-in token and user management with expiration, quotas, IP restrictions, and model allowlists.
• Supports multi-machine master/slave deployment for scaling beyond a single node.
• Customizable web UI (logo, themes, pages) and a management API for external integrations.

Caveats

• Model mapping rewrites the request body rather than passing it through, which the README warns can silently drop fields the project does not yet officially support.
• The default root password is 123456, and the README explicitly warns you to change it immediately after first login.
• WeChat OAuth login requires deploying an additional companion service.

Verdict

Worth a look if you are running a multi-tenant service, an internal AI platform, or any setup where you need to ration and route access across many LLM vendors through one familiar API surface. If you only ever call OpenAI directly, it is overkill.