Zendriver-MCP

A powerful MCP (Model Context Protocol) server for browser automation using Zendriver - an undetectable, async-first browser automation framework. Built specifically for LLM-powered automation with a focus on token efficiency.

Why Zendriver?

Most browser automation tools (Selenium, Playwright, Puppeteer) use WebDriver protocol, which is easily detected by anti-bot systems. Zendriver is different:

Use cases where Zendriver is better than others:

• Automating sites with bot protection (Cloudflare, Akamai, etc.)
• Scraping dynamic SPAs that block traditional automation
• Testing authenticated workflows on protected sites
• Building AI agents that interact with the real web

Features

• Undetectable - Uses Chrome DevTools Protocol, bypassing WebDriver detection
• Token-Optimized DOM Walker - 78% reduction in token usage vs raw HTML
• 35+ Essential Tools - Focused, powerful browser automation capabilities
• Modern Web Support - Works with contenteditable divs, SPAs, and dynamic content
• Smart Element Handling - Auto-skips hidden elements, provides selector suggestions
• CDP Network Logging - Real-time network request and console log capture, its super easy to create endpoint based scrappers as llms can directly access the network logs
• Security Auditing - Comprehensive security analysis tool

Usage with Claude Desktop

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "zendriver": {
      "command": "python",
      "args": ["path/to/Zendriver-MCP/run.py"]
    }
  }
}

Token Optimization Protocol

The Problem

Traditional approaches send raw HTML or verbose element trees to LLMs:

The Solution

Our DOM walker produces compact, semantic output:

{"id": 1, "t": "btn", "l": "Search", "r": "hdr"}

Optimization Techniques

Real-World Results

Perplexity.ai homepage:

• Raw HTML: ~45KB (~11,000 tokens)
• Standard element dump: 95 elements (~2,800 tokens)
• Our optimized output: 17 elements (~400 tokens)
• Total reduction: 96% fewer tokens

Label Inference Priority

Instead of showing “(unlabeled)”, we infer labels from multiple sources:

1. aria-label attribute
2. aria-labelledby reference
3. Associated `` element
4. placeholder attribute
5. Direct text content
6. title attribute
7. alt attribute (for images)

Region Detection

Elements are tagged with their page region for context:

• hdr - Header/banner area
• nav - Navigation
• main - Main content
• side - Sidebar/aside
• ftr - Footer
• dlg - Modal/dialog

Usage

# Get the optimized interaction tree
tree = get_interaction_tree()
# Returns: [{"id": 1, "t": "btn", "l": "Submit", "r": "main"}, ...]

# Click using the numeric ID
click("1")  # Clicks the element with id=1

# Type into an input
type_text("Hello", "3")  # Types into element with id=3

This approach lets LLMs work with web pages using minimal context while maintaining full functionality.

License

MIT