UI-TARS Browser MCP

ByteDance's browser automation MCP server, from its UI-TARS Desktop repository and published on npm as @agent-infra/mcp-server-browser. It drives a browser with Puppeteer and gives your agent the page as structured accessibility data and numbered clickable elements rather than pixels, so it works without a vision model. An optional vision mode adds screenshots and coordinate clicks for complex layouts. What you can do with it - Open pages, go back and forward, and open, switch and close tabs - List the clickable elements on a page, then click, hover, select options and fill in form fields - Press keys and scroll - Read a page as text or Markdown, and collect its links - Take screenshots of the whole page or a single element - Run JavaScript in the page - See downloaded files and the browser's console logs How it connects It runs on your machine with npx @agent-infra/mcp-server-browser@latest and needs Node.js 18 or newer. ByteDance documents setup for VS Code with GitHub Copilot, Cursor, Windsurf and Claude Desktop. It opens a visible browser by default: pass --headless to hide it, --browser to pick Chrome, Edge or Firefox, or --cdp-endpoint or --ws-endpoint to drive a browser that's already running. Start it with --port to serve it over Streamable HTTP or SSE instead of stdio. Good to know - Add --vision to work from screenshots and coordinate clicks instead of accessibility snapshots. - Other options set a proxy, user agent, viewport size, user data directory and output folder. - It needs no account or API key, but it acts in the browser with whatever sessions that browser holds, and its browser_evaluate tool runs any JavaScript on the page.