Agent-Browser MCP Server

Overview

agent-browser is a fast, native Rust CLI for browser automation via CDP with accessibility-tree snapshots. Unlike traditional browser automation tools that rely on visual screenshots, agent-browser uses accessibility tree data for precise, efficient browser interaction.

With over 11,753 GitHub stars, agent-browser has become a popular choice for AI agent browser automation, especially for accessibility-focused workflows and fast browser-based tooling.

Features

• Accessibility-tree snapshots — Precise element detection via accessibility tree
• Native Rust implementation — Fast, memory-efficient browser automation
• CDP protocol support — Works with Chrome, Chromium, and Edge
• MCP integration — Standard Model Context Protocol support
• Cross-platform — Supports Windows, macOS, and Linux
• Zero dependencies — No heavy browser automation frameworks required
• Real-time interaction — Low-latency browser commands
• Headless mode — Fully headless execution for CI/CD

Installation

Quick Setup

# Install via npm (recommended)
npm install -g @anthropic-ai/agent-browser

# Or install from source
cargo install agent-browser

Claude Desktop Configuration

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "agent-browser": {
      "command": "agent-browser",
      "args": ["serve"],
      "env": {
        "AGENT_BROWSER_URL": "http://localhost:3000"
      }
    }
  }
}

Docker Setup

docker run -p 3000:3000 agent-browser:latest

Setup

1. Install Chrome or Chromium

   • Ensure Chrome/Chromium is installed and accessible
   • For headless mode, ensure Chrome can run headless

2. Start the Browser

   agent-browser start
   

3. Connect to MCP

   agent-browser serve
   

Available Tools

Tool	Description
browser_navigate	Navigate to a URL
browser_snapshot	Get accessibility tree snapshot
browser_click	Click an element by selector
browser_type	Type text into an element
browser_scroll	Scroll the page
browser_back	Navigate back
browser_forward	Navigate forward
browser_refresh	Refresh the page
browser_get_text	Get text content of element
browser_get_html	Get HTML content of element
browser_take_screenshot	Take a screenshot
browser_execute_js	Execute JavaScript

Usage Examples

Navigate to a URL

{
  "name": "browser_navigate",
  "arguments": {
    "url": "https://github.com/anthropics/agent-browser"
  }
}

Get Accessibility Snapshot

{
  "name": "browser_snapshot",
  "arguments": {
    "depth": 5
  }
}

Click an Element

{
  "name": "browser_click",
  "arguments": {
    "selector": "#submit-button"
  }
}

Type Text

{
  "name": "browser_type",
  "arguments": {
    "selector": "#search-input",
    "text": "MCP servers",
    "slowly": true
  }
}

Take Screenshot

{
  "name": "browser_take_screenshot",
  "arguments": {
    "fullPage": false,
    "format": "png"
  }
}

Advanced Features

Accessibility Tree Parsing

agent-browser excels at accessibility tree parsing, providing structured data about page elements:

{
  "name": "browser_snapshot",
  "arguments": {
    "includeHidden": false,
    "maxDepth": 10
  }
}

Headless Mode

For CI/CD and automated workflows:

agent-browser start --headless

Multi-Tab Support

agent-browser supports multiple browser tabs:

{
  "name": "browser_new_tab",
  "arguments": {
    "url": "https://example.com"
  }
}

Form Filling

Automated form filling with accessibility-aware selectors:

{
  "name": "browser_fill_form",
  "arguments": {
    "fields": {
      "username": "testuser",
      "password": "securepassword123"
    },
    "submit": true
  }
}

Integration with AI Agents

Claude Code Integration

agent-browser works seamlessly with Claude Code for browser automation:

# In Claude Code, use browser tools
> Click the login button
> Type my email in the input
> Submit the form

Agent Workflow Example

# Example agent workflow using agent-browser
async def search_and_extract(url: str, query: str):
    # Navigate
    await browser_navigate(url)
    
    # Search
    await browser_type("#search", query)
    await browser_click("#search-button")
    
    # Extract results
    snapshot = await browser_snapshot()
    return parse_results(snapshot)

Pros

• ✅ Fast native Rust implementation
• ✅ Accessibility-tree snapshots for precise interaction
• ✅ MCP standard integration
• ✅ Cross-platform support
• ✅ Low latency browser commands
• ✅ Headless mode for CI/CD
• ✅ Works with any CDP-compatible browser

Cons

• ❌ Requires Chrome/Chromium installation
• ❌ Accessibility tree may miss visual-only elements
• ❌ Less mature than Playwright/Puppeteer
• ❌ No built-in waiting mechanisms
• ❌ Limited screenshot-based debugging

When to Use

• AI agent browser automation
• Accessibility-focused web scraping
• Fast browser-based tooling
• MCP-integrated workflows
• Headless browser automation
• Form filling and submission

Resources

• agent-browser GitHub
• CDP Protocol Documentation
• MCP Documentation
• Accessibility Tree Guide