zendriver-mcp

Undetectable browser automation for LLM agents, spoken over MCP.

zendriver-mcp is an MCP server that gives your coding agent (Claude, Cursor, Gemini, Copilot) a real Chrome browser it can actually use on the real web - behind Cloudflare, behind login walls, on pages that detect and block WebDriver.

It's built on Zendriver, which speaks the Chrome DevTools Protocol directly instead of going through WebDriver. No navigator.webdriver flag, no headless telltales, a fingerprint that looks like ordinary Chrome.

On that foundation, zendriver-mcp layers everything an agent needs to get work done: a token-efficient DOM walker, an accessibility tree with stable uids, performance traces, Lighthouse audits, heap snapshots, human-like input, device emulation, cookie round-tripping, request interception, and more - 96 tools across 22 modules.

Why

Most browser-automation MCP servers lean on Puppeteer or Playwright, which means WebDriver, which means navigator.webdriver === true and a shopping list of headless telltales Cloudflare / Akamai / PerimeterX know by heart. Agents that try to work on the real web hit bot walls immediately.

zendriver-mcp starts from the opposite direction:

• CDP instead of WebDriver, so the fingerprint stays clean.
• Human-like input primitives (bezier mouse paths, gaussian typing) for sites that look at how you click, not just whether you do.
• Cookie round-tripping so you log in once and reuse the session.
• A Cloudflare Turnstile solver for the last-mile interactive challenge.

Quick start

Getting started Tool reference

Published on PyPI and the MCP Registry as io.github.bituq/zendriver-mcp.

uvx zendriver-mcp            # zero-setup, re-resolves every run
uv tool install zendriver-mcp  # install once, invoke many
pipx install zendriver-mcp
pip install zendriver-mcp

Hook it up to Claude Desktop / Code in one JSON block:

{
  "mcpServers": {
    "zendriver": {
      "command": "uvx",
      "args": ["zendriver-mcp"]
    }
  }
}

Highlights

StealthDevToolsSessionObservability

• bypass_cloudflare solves Turnstile challenges
• set_user_agent, set_locale, set_timezone, set_geolocation
• human_click uses bezier mouse paths
• human_type uses gaussian inter-keystroke timing

• start_trace / stop_trace produce DevTools-loadable JSON
• take_heap_snapshot produces standard .heapsnapshot
• run_lighthouse reuses the already-running browser
• set_cpu_throttle, set_network_conditions, set_viewport

• export_cookies / import_cookies work on HTTP-only cookies
• block_urls, set_extra_headers, bypass_service_worker
• grant_permissions, reset_permissions
• configure_proxy restarts with a --proxy-server arg

• get_interaction_tree - 96% fewer tokens than raw HTML
• get_accessibility_snapshot - stable uids that survive re-renders
• start_screencast -> export_screencast_mp4 via ffmpeg
• mock_response / fail_requests via Fetch.enable

Recipes

Step-by-step walkthroughs for common workflows:

• Scrape behind a login, once
• Beat Cloudflare Turnstile
• Performance audit an authenticated flow
• Mock an API response to test the UI

License

MIT. The original 49-tool foundation and the token-optimised DOM walker come from ShubhamChoulkar/Zendriver-MCP.