Gemini load balancer that speaks fluent OpenAI

What it does

Gemini Balance is a FastAPI-based proxy that sits between your applications and Google’s Gemini API. It manages a pool of API keys, distributing requests sequentially while translating incoming calls into either native Gemini or OpenAI-compatible formats. The app also provides a web dashboard for real-time key monitoring and visual configuration changes that apply without restarting the server.

The interesting bit

The project doesn’t just forward traffic; it adapts Gemini’s native capabilities—like imagen-3.0-generate-002 image generation and web search—into OpenAI-style endpoints, even offering “fake streaming” through its Hugging Face-compatible route. It also auto-disables exhausted or failing keys and periodically resurrects them, effectively treating API keys like cattle rather than pets.

Key highlights

• Sequential polling across multiple Gemini API keys with automatic retry and failure-based disabling.
• Dual-protocol support: exposes both native Gemini (/gemini/v1beta) and OpenAI-compatible (/hf/v1, /openai/v1) endpoints for chat, embeddings, and image generation.
• Visual admin panel for key management, model filtering, and live configuration updates without restarts.
• Integrated image hosting uploads to SM.MS, PicGo, Cloudflare ImgBed, or Aliyun OSS for generated images.
• Optional web search and image-editing capabilities invoked by suffixing model names with -search or -image.

Caveats

• Licensed under CC BY-NC 4.0: commercial resale is explicitly prohibited, and the author warns against third-party resellers.
• The “advanced features” route advertises “fake streaming,” which may or may not match your expectations for real-time token delivery.
• The extensive configuration surface—multiple upload providers, thinking budgets, URL context—means setup can get dense before the first request goes through.

Verdict

Ideal for developers running multiple Gemini free-tier or paid keys who want to unify them behind a single OpenAI-compatible URL. Less useful if you already have a single, high-rate-limit enterprise key and no need for protocol translation.