A Gemini proxy that speaks OpenAI and fakes its own streams

What it does

This FastAPI proxy translates OpenAI-format requests into Google Gemini calls. It exposes standard /v1/chat/completions and /v1/models endpoints, rotates through multiple API keys, and enforces rate limits and optional password protection. The README targets free-tier hosts like Hugging Face Spaces and Claw Cloud, but it runs anywhere.

The interesting bit

Its default “fake streaming” mode sends empty keep-alive chunks to the client while it fetches a full non-streaming response from Gemini in the background, then flushes the result in one go. There is also an optional “disguise” mode that appends random noise strings to prompts in an attempt to avoid automation detection—though the README warns this confuses non-SillyTavern clients unless disabled.

Key highlights

• Full OpenAI API compatibility for chat completions, function calling, and model listing, handling text, image, and file inputs.
• API key polling with per-key daily quotas, per-IP daily caps, and per-minute rate limits.
• Optional concurrency cache: fire multiple identical requests simultaneously and bank the extras for instant regenerations (exact-match only, and only with non-streaming or fake-streaming).
• Search mode via -search suffixed models to give Gemini live web access.
• Supports both Vertex AI and AI Studio credential modes.

Caveats

• The default PASSWORD is 123; the README explicitly warns you to change it before production.
• Fake streaming is enabled by default, so real token-by-token streaming requires explicitly turning the feature off.
• The random-string disguise is tuned for SillyTavern and will break clients like Cherry Studio unless turned off.

Verdict

Useful if you want a quick, drop-in OpenAI facade for Gemini on free hosting. Less appealing if you already have clean Gemini connectivity and no need for proxy indirection or compatibility shims.