zexadev/gemini-web2api-go

gemini‑web2api‑go – Gemini Web ↔ OpenAI‑compatible API proxy

What it is

• A single‑binary Go server that reverse‑proxies the private browser‑based protocol used by Google Gemini’s web UI and re‑exposes it as a OpenAI‑compatible REST API (/v1/chat/completions, /v1/models, etc.).
• Because it talks to the web UI instead of the official Google Generative Language API, no Google API key or paid quota is required – the server can operate anonymously or with a logged‑in Google account (via cookies) to unlock extra Gemini features.

Key features

Feature	Details
OpenAI‑compatible endpoints	/v1/chat/completions, /v1/models, /v1/responses, plus a /v1/videos endpoint that mimics OpenAI Sora’s async video generation.
Model support	gemini-3.6-flash, gemini-3.5-flash-lite (anonymous). With a valid Google cookie you also get gemini-3.1-pro, extended‑thinking variants, image/audio generation (gemini-image, gemini-music, gemini-canvas, gemini-video).
True streaming	For normal chat the server forwards each upstream frame as a delta chunk, giving low‑latency streaming. Tool‑calling requests are buffered and sent after the full response (required by Gemini).
Authentication	Bearer token or x‑api‑key (generated on first start, rotatable via the admin UI).
Rate‑limiting & proxy pool	Per‑IP concurrency, RPM and RPH limits; configurable proxy pool with automatic failure‑circuit‑breakers; each proxy gets its own quota slot.
Cookie pool	Stores multiple Google login cookies, rotates them per‑request, auto‑renews cookies every 10 min, and binds each cookie to a single outbound IP to avoid Google‑detected account sharing.
Persistence	SQLite by default (30‑day request logs, permanent aggregates). MySQL/PostgreSQL can be used by setting SQL_DSN.
Admin UI	A lightweight Chinese‑language web panel (/admin) for monitoring KPIs, viewing request metadata, managing the proxy and cookie pools, and changing runtime configuration without restarting.
MCP (web‑search) server	Exposes Gemini’s built‑in web‑search tool at /mcp so clients like Claude Desktop can call web_search(query) and receive a synthesized answer with source links.
Zero‑code deployment	Single binary for six platforms, Docker distroless image, and a Docker‑Compose starter.
Privacy‑first	Prompt and response bodies are never stored; only metadata (model, token counts, latency, status) is persisted.

How to run

1. Binary – download the appropriate release, make it executable and start it:

   chmod +x gemini-web2api-go_*
   ./gemini-web2api-go_* --port 8083 --admin-token mytoken
   

2. Docker – one‑liner:

   docker run -d --name gemini-web2api \
     -p 127.0.0.1:8083:8083 \
     -v "$PWD/data:/data" \
     -e ADMIN_TOKEN=mytoken \
     ghcr.io/zexadev/gemini-web2api-go:latest
   

3. Source – go build -o gemini-web2api-go . and run the same flags.

Calling the API (any OpenAI‑compatible client works):

curl http://localhost:8083/v1/chat/completions \
  -H "Authorization: Bearer sk-gemini-..." \
  -H "Content-Type: application/json" \
  -d '{"model":"gemini-3.6-flash","messages":[{"role":"user","content":"Hello!"}]}'

Or with the Python SDK:

from openai import OpenAI
client = OpenAI(base_url="http://localhost:8083/v1", api_key="sk-gemini-...")
resp = client.chat.completions.create(model="gemini-3.6-flash",
                                   messages=[{"role":"user","content":"Explain quantum entanglement"}])
print(resp.choices[0].message.content)

When to add a Google cookie

• Anonymous mode only gives the two flash models and the built‑in web‑search tool.
• Adding a SID/HSID/... cookie (via the admin UI or --cookie-file) unlocks:
  • gemini-3.1-pro with reasoning chain (reasoning_content)
  • “thinking” extensions for the flash models
  • multimodal generation (images, music, canvas, video)
  • longer effective context (≈ 160 KB vs 130 KB) via automatic attachment of a message.txt file.

Limitations / gotchas

• The server only implements the OpenAI‑style endpoints; Gemini’s native /v1beta endpoints are not supported.
• Token limits are enforced on the raw UTF‑8 byte size of the request (default 128 KB). Exceeding it returns context_length_exceeded for anonymous users; with a cookie the request is turned into a text attachment up to ≈ 160 KB.
• Streaming works for normal chat; tool‑calling and /v1/responses are buffered because Gemini needs the full tool output before it can continue.
• Proxy‑pool traffic uses Go’s standard library TLS fingerprint (not the Chrome 146 fingerprint used for direct connections). This does not materially affect capacity but is worth noting.
• Image/audio generation counts only the textual description toward completion_tokens; the base64 payload is excluded.
• The admin UI is Chinese‑only; the underlying API is language‑agnostic.

Typical use cases

• Run locally‑hosted LLM‑aware applications (ChatGPT‑style UI, LangChain agents, code assistants) without needing a Google Cloud billing account.
• Provide a drop‑in OpenAI endpoint for tools that only understand the OpenAI schema (e.g., newapi, Cherry Studio, Cursor, dify).
• Enable Gemini’s web‑search tool for agents that support “MCP” (Claude Desktop/Code) via the /mcp endpoint.
• Experiment with Gemini’s multimodal models (image, music, canvas, video) behind a familiar OpenAI API surface.

---

TL;DR: gemini-web2api-go is a Go‑based reverse‑proxy that turns the private Gemini web UI protocol into a fully‑featured, OpenAI‑compatible API, with optional Google‑cookie support for premium Gemini models, built‑in rate‑limiting, proxy & cookie pools, and a tiny admin dashboard. It runs as a single binary or a distroless Docker container and requires no Google API key.