claude-drives-gemini

English | 简体中文 | 日本語 | 한국어

A Claude Code skill: let your AI agent drive the Chrome where your Google AI Pro account is already signed in, and generate images (Nano Banana, 2816×1536) / video (Veo) / music (Lyria) through the Gemini web app. Output lands on disk losslessly, and media bytes never enter the agent's context.

Make the Google Gemini subscription you already pay for do extra work.

Features

  • One command, end to end: auto-launch the dedicated Chrome → ensure Pro + Extended Thinking → submit the prompt → stage-aware wait → native download → optional watermark removal, returning a single line of JSON.
  • Never burns quota twice: only an explicit refusal triggers a safe resubmit; a timeout only extends the wait; a failed download only retries the download.
  • Full resolution, lossless: the native 2816×1536 original is downloaded directly; --dewatermark removes the visible corner logo with exact math (SynthID excluded — see the risk section).

Prerequisites

  • Node ≥ 22 + Chrome + a Google AI Pro (or higher) subscription — the consumer web subscription, not an API key.
  • Verified on Windows; the macOS / Linux scaffolding is in place but untested (feedback welcome).
  • A local proxy if your network cannot reach Google directly.
  • UI text matching is verified against the Chinese Gemini interface; locales/en.json is an unverified best guess — calibration reports are very welcome.

Quick start

Recommended: let Claude Code install it for you. Paste this to Claude Code:

Install claude-drives-gemini for me: clone https://github.com/MagicYangG/claude-drives-gemini into ~/.claude/skills/claude-drives-gemini, run node setup/init.mjs to generate the local config, walk me through the one-time seeding login, then run the start-chrome launcher and verify with node scripts/smoke.mjs.

Manual install is three steps:

git clone https://github.com/MagicYangG/claude-drives-gemini ~/.claude/skills/claude-drives-gemini
cd ~/.claude/skills/claude-drives-gemini && node setup/init.mjs   # detect Chrome, write config, generate launchers
  1. Seeding login (one time): run launchers/login-chrome.cmd|sh, sign in to Google in the clean window that opens, then close it. (You must sign in from this window — Google refuses logins in a window that exposes a debugging port.)
  2. Start the automation Chrome: run launchers/start-chrome.cmd|sh.
  3. Verify: node scripts/smoke.mjs (zero-quota smoke test: environment → Chrome → open page → probe controls).
  4. Create: just tell Claude "generate an image with Gemini…", or run one command:
node scripts/gemini-gen.mjs image "Generate an image directly: a stone bridge in a riverside town at dusk after rain, watercolor style. No explanation." out/image.png --dewatermark

Video and music work the same way — swap image for video / music. The prompt must contain an explicit generation verb ("generate/create … directly" + "no explanation"), otherwise Gemini misroutes it. Add --use-tool for 9:16 vertical video or the built-in style templates. --dewatermark needs a one-time helper install — node scripts/setup-gwr.mjs — otherwise the flag is skipped silently (the result JSON reports dewm:false). Step-by-step command chains and the full script reference live in scripts/README.md.

⚠️ Risks and disclaimer (please read)

  • This project is not affiliated with Google. Automating the consumer Gemini web app may violate Google's Terms of Service, and your account (especially a paid subscription) could be restricted or banned. Use at your own risk; consider the official API as an alternative.
  • Built-in anti-abuse discipline: serial rather than parallel, throttled output, real keyboard/mouse events, quota respected — please do not turn it into a concurrent flood.
  • Watermarks: only the visible corner logo is removed (reverse alpha blending, mathematically lossless). The SynthID provenance watermark is not removed and cannot be, so generated content remains traceable.

Layout

SKILL.md            # Agent entry point: ~4KB core contract (progressive disclosure)
setup/init.mjs      # First-run wizard: detect Chrome → write config → generate local launchers
launchers/          # Launcher templates (*.tpl; init generates local start/login-chrome.cmd|sh)
scripts/            # All executable scripts + the operations manual README.md
references/         # runtime-env (environment/login/anti-ban) · gemini-ui (living selector archive) · quota-tiers · dewatermark
locales/            # UI text matching (zh-CN verified; en pending calibration)
providers/          # ADR for the v2.0 provider refactor

License

MIT