Hijacking Gemini’s web UI for agents that need pictures

What it does gemini-skill is an MCP server that drives the Gemini web interface via Chrome DevTools Protocol. It exposes image generation, multi-turn chat, reference-image uploads, and conversation-asset extraction as standard MCP tools for any compatible client.

The interesting bit Rather than calling a REST endpoint, it launches a stealth-hardened browser, navigates to gemini.google.com, and literally clicks the buttons a human would. A background daemon keeps the browser alive for 30 minutes of idle time, so cold-start latency is amortized across calls.

Key highlights

• Wraps Gemini web features—model switching, thinking-depth controls, and watermark-free image downloads—as MCP primitives
• Uses puppeteer-extra-plugin-stealth to hide automation fingerprints from Google’s front end
• Gracefully adapts to UI changes: selectors carry fallbacks for both pre- and post-May 2026 Gemini layouts
• Can piggyback on an existing OpenClaw browser session via a shared CDP port, though this disables stealth hardening
• Bundles a secondary Atlas Cloud provider for conventional OpenAI-compatible API calls when browser automation is overkill

Caveats

• Requires a Google account signed in inside the automated browser; headless mode is off by default specifically to let you log in manually
• Reusing an OpenClaw browser session means losing the stealth plugin’s anti-detection protections, which the README warns is more likely to trigger automation blocks

Verdict Useful if you need to wire Gemini’s web-tier capabilities into an agent pipeline today. Avoid if you require a supported, first-party API rather than a browser-puppeting workaround.