Releases · Lutiancheng1/gemini-webapi-proxy

Highlights

Curated model list — 2 chat (gemini-3-flash, gemini-3-pro) and 2 image aliases (gemini-2.5-flash-image, gemini-2.5-pro-image).
OpenAI-compatible — POST /openai/v1/chat/completions and POST /openai/v1/images/generations both verified end-to-end against gemini-3-flash, gemini-3-pro, and both image models.
Pluggable cookie sources — desktop browser (Safari/Chrome/…), env vars, Netscape cookies.txt.
Pluggable image downloader chain — Playwright Chromium → httpx → curl_cffi → library save.
Upstream refusal detection — chat refusals like "I cannot fulfill this request." are mapped to HTTP 403, not silently 200.
Bilingual docs — README + docs/ in English and 中文.

ModelRegistry.resolve_id() no longer back-matches stable image aliases, so asking for gemini-3-pro can't accidentally route to an image alias when the underlying model is missing from _runtime (e.g. transient UNAUTHENTICATED).
Dockerfile switched from a builder-stage wheel to pip install -e . of the live source tree, so a host-side code change always invalidates the build.
apt-get install wrapped in a 3-attempt retry loop to absorb transient 503s from deb.debian.org on the GitHub runner.

pip install git+https://github.com/Lutiancheng1/gemini-webapi-proxy.git@v0.1.0

Or with the browser-cookie extra (Safari/Chrome/... read on every request):

pip install "gemini-webapi-proxy[browser-cookie] @ git+https://github.com/Lutiancheng1/gemini-webapi-proxy.git@v0.1.0"

See docs/docker.md for the full Docker guide, and docs/studio.md for connecting Outsider Studio.