What is gemini-web2api-go?

gemini-web2api-go reverse-proxies the Gemini web application (gemini.google.com) into an OpenAI-compatible REST API. It requires no Google API key and works with free Gemini accounts. The project is a single Go binary (or distroless Docker image) that implements Chrome 146 TLS fingerprinting, automatic cookie renewal, a proxy pool with per-IP rate limiting, and a Chinese-language admin dashboard for monitoring and configuration. It exposes /v1/chat/completions (streaming), /v1/models, /v1/responses, and /v1/videos for async video generation. An MCP server at /mcp provides a web_search tool for Claude Desktop and Cursor. Supported models include gemini-3.6-flash, gemini-3.5-flash-lite (anonymous), and gemini-3.1-pro, gemini-3.8-flash, gemini-4-argon with thinking variants when a valid cookie is provided. Image (gemini-image), music (gemini-music), canvas (gemini-canvas), and video (gemini-video) generation are supported. Data persists in SQLite by default, with optional MySQL/PostgreSQL via SQL_DSN. Deployment options: Docker (ghcr.io/zexadev/gemini-web2api-go), Docker Compose, direct binary download, or building from source. The admin UI at /admin manages proxy/cookie pools, API keys, and runtime settings without restart. Known limits: per-IP request caps (80–180 before Google blocks), context length ~130 KB (160 KB with cookie), prompt-level function calling, and Chrome-exported cookies may expire quickly due to device-bound sessions (Firefox recommended).

Should you self-host gemini-web2api-go?

gemini-web2api-go scores 56/100 on selfhost.directory (how scoring works) and ranks #14 of 39 in API Gateway. It has 722 GitHub stars.

It shipped 10 stable releases in the past year. The latest, v4.26.0, came out on 26 October 2026. New versions land roughly every 4 days.

Setup is moderate: expect some configuration beyond a single command. An official Docker image is available. It is MIT-licensed and written mostly in Go.

Its closest alternative on selfhost.directory is Kong Gateway (82/100, 44 314 stars). It scores higher overall, so compare the two before you pick. See gemini-web2api-go vs Kong Gateway side by side.

Features

Admin Dashboard
Canvas generation
Chrome TLS fingerprintingProxy connections use Go stdlib TLS fingerprint
Cookie pool with auto-renewalFirefox-exported cookies recommended for long-term renewal
Function calling (prompt-level)Model may not always return formatted tool calls; reliable for private data queries
Image GenerationResolution locked to ~1MP by upstream
MCP web_search tool
Multiple Database BackendsAuto-migration, shared schema
Multi-turn conversation supportDoes not expand model context window
Music Generation
OpenAI Compatible API
Proxy pool with per-IP rate limiting
Reasoning contentreasoning_tokens counted separately from completion_tokens
Streaming ResponsesTools requests and /v1/responses are buffered
Video GenerationFree accounts rejected by upstream content policy

FAQ

Releases 里挑对应平台的下载，不用 Go、不用 Docker，单个文件就是全部

Pre-built binaries for 6 platforms available on GitHub Releases. No Go or Docker needed — single file is everything.

支持 MySQL / PostgreSQL 吗？

Yes. Set SQL_DSN env var (mysql://... or postgres://...). Defaults to SQLite. Auto-migration, shared schema across all three.

为什么在 Dify / 反向代理后面流式变成了一次性输出？

Regular chat is true streaming. Tools requests and /v1/responses are buffered. Response includes X-Accel-Buffering: no; check your reverse proxy's buffering settings.

生图用的是哪个模型？能出 2K 吗？

Uses whatever image backend the web UI currently uses (e.g., Nano Banana 2). Resolution locked to ~1 megapixel, no 2K output.

能用 Gemini 4 Argon 吗？怎么用？

Yes, model name gemini-4-argon (and -thinking, -thinking-medium). Requires a cookie from an account that has 4 Argon visible in gemini.google.com model selector (Google AI Pro/Ultra).

有 gemini-3.8-flash 吗？

Yes (gemini-3.8-flash and gemini-3.8-flash-thinking), requires paid Gemini account. Free accounts get downgraded to 3.5 Flash-Lite. gemini-3.7-flash kept as alias.

能当 Cursor / Dify / WorkBuddy 这类 agent 工具的后端吗？

Yes, OpenAI-compatible. Works for chat and coding. But context limit is ~130KB (160KB with cookie), so large agent prompts may hit limits. Function calling is prompt-level, not protocol-level.

为什么每发一条消息，gemini.google.com 里就多出一个会话？

Each request is an independent session by default. Enable 'multi_turn' in settings to reuse a conversation_id. Or enable 'temporary_chat' (recommended) to keep sessions out of the web UI entirely.

免费账号和 Pro / 付费账号差在哪？

Anonymous: gemini-3.6-flash, gemini-3.5-flash-lite with web search. Free account cookie adds: gemini-3.1-pro, thinking variants, vision, image/music/canvas. Paid adds: gemini-3.8-flash, gemini-4-argon, video generation.

会被封号吗？封禁是按账号还是按 IP？

Bans are by exit IP + TLS fingerprint, not by account. Burst requests (80-180) trigger blocks; slow pacing (10/min) can run 800+ requests unblocked. Paid accounts don't raise limits.

和 Sophomoresty/gemini-web2api（Python）、其它叫 gemini-web2api 的 Go 项目是什么关系？

Independent Go implementation. Early protocol research referenced the Python version's approach, but it's not a port. Other same-named Go projects are unrelated.

这和 Google 官方 Gemini API 有什么区别？需要 API Key 吗？

No API key needed. This project reverse-proxies the gemini.google.com web interface directly, not the official generativelanguage.googleapis.com API. So there's no API key or paid quota.