I built barebrowse — an open-source MCP server that gives AI agents a real browser.
No Playwright. No Selenium. No bundled Chromium. No webMCP, No 200MB download. 11 modules, 2,400 lines, zero dependencies.
Today I asked Claude to find 5 electronic components from a single Dutch webshop. It hit the usual wall — prices rendered client-side, invisible to every fetch tool. So it opened barebrowse, browsed 7 product pages, read JS-rendered prices and stock levels, compared board variants, and handed me a complete cart. Under 2 minutes.
What makes it different:
- Uses the browser you already have — your sessions, your cookies, your logins
- Prunes pages down to what matters. 40-90% fewer tokens than raw ARIA output
- Handles cookie consent in 29 languages, bot detection, SPAs, permission prompts — the agent doesn't think about any of it
- 3 modes: headless (fast), headed (your real browser), hybrid (tries headless, falls back if blocked)
- Works with Claude Code, Cursor, VS Code, Claude Desktop — anything that speaks MCP
One command:
npx barebrowse mcp
12 tools: goto, snapshot, click, type, scroll, drag, upload, pdf, and more. The agent browses like you would.
Tested against 20+ sites across 8 countries. Google, GitHub, Amazon, BBC, Reddit, Wikipedia — all consent walls dismissed, all interactions working.
Part of the bare ecosystem: bareagent (think→act loop), barebrowse (web), baremobile (Android). All vanilla JS, all zero dependencies, all MIT.
https://github.com/hamr0/ba...