Automates the Google Gemini web app using a signed-in Google account, enabling prompt submission, chat continuation, file attachments, model and tool selection, chat listing, file downloads, and general browser page interaction.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@gemini-mcp summarize the latest AI news in five bullet points"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
An MCP server that lets Claude (or any MCP client) use the Google Gemini web app with your normal Google account. No API key.
It drives the real Gemini website in its own Chrome window. The tools find buttons, links and text boxes by their accessible role and name (what a screen reader sees), not by CSS selectors, so the agent can re-read the page and keep working when Google changes the layout.
Unofficial. This is not affiliated with Google. It automates the normal website, which may go against Google's Terms of Service and may stop working if Google blocks automated browsers. Use it for personal and experimental work, at your own risk. If you need something supported, use the paid Gemini API instead.
Shortcuts:
Tool | What it does |
| Send a prompt and return the finished reply as markdown. Options: continue a chat ( |
| List recent chats with their URLs |
| Click a download button (e.g. "Download full size image") and save the file (default |
General tools, which work on any page:
Tool | What it does |
| Accessibility tree of the page: every element's role, name and text |
| Go to a URL (default: a new chat) |
| Click an element by role/name, or by visible text |
| Fill a text field and optionally press Enter |
| Press a key, e.g. |
| Click an upload button and give it local files |
| Read the visible text of the page or one element |
| Wait for a reply to finish, or until some text appears |
| PNG of the page when the text view is not enough |
Related MCP server: WebCore-MCP
Needs Node.js 18+ and a Chrome-type browser. It uses, in order: GEM_CHROME, an installed Google
Chrome or Chromium, or a Chromium downloaded by Playwright (npx playwright install chromium).
git clone https://github.com/DwytMCL/gemini-mcp.git
cd gemini-mcp
npm install
node server.js login
npm test # starts the server, lists the tools and opens Geminilogin opens a separate Chrome window with its own profile. Sign in, check Gemini loads, then
quit that window (Cmd+Q). The server cannot attach while the sign-in window is still open.
Sign in separately for each tool. Never copy a signed-in profile folder to another browser or server: Google sees the same session in two browsers, revokes it, and both come back signed out after the next restart.
claude mcp add -s user gemini -- node /absolute/path/to/gemini-mcp/server.jsAdd to claude_desktop_config.json and restart Claude Desktop:
{
"mcpServers": {
"gemini": { "command": "node", "args": ["/absolute/path/to/gemini-mcp/server.js"] }
}
}All optional, set as environment variables:
Variable | Default | Meaning |
| Google Chrome, else Playwright's Chromium | Path to the Chrome/Chromium binary |
|
| Browser profile holding your Google sign-in |
|
| Local port the server uses to control Chrome |
|
| Where |
Your Google session lives only in the browser profile folder on your machine (GEM_PROFILE).
Nothing is sent anywhere except to Google by the browser itself. Treat that folder like a password:
don't share or commit it. Chrome's control port listens on 127.0.0.1 only, but any program on your
machine could use it while the window is open.
MIT
This server cannot be deployed
Hosted real Google Chrome MCP with per-user persistent state. Navigate, click, type, screenshot.
Give AI agents a real cloud Chrome: open sessions, navigate, read pages, click, type and screenshot.
Browser MCP for logged-in tasks. Uses your Chrome — credentials stay local. Zero-token replay.
Real Chrome for agents: start a browser, read pages as numbered markdown, click, type, hand off.
