AgenticBrowser

Fully autonomous browser runtime for AI agents. Zero human intervention. If a human can access it, the agent accesses it.

What It Does

AgenticBrowser gives AI agents a real browser that handles the modern web autonomously:

• Opens any URL — JavaScript SPAs, iframes, shadow DOM, lazy loading, all handled
• Auto-solves challenges — Cloudflare, reCAPTCHA, hCaptcha resolved automatically
• Reads clean content — Smart multi-source extraction to markdown
• Interacts by intent — "click the login button", "type in the search box"
• Extracts structured data — Schema-driven extraction from any page
• Verifies goals — "user is logged in", "page contains pricing"
• Recovers from blocks — Tries reader mode, print version, viewport switching, etc.

How It's Different

Feature	Playwright	gsd-browser	AgenticBrowser
Anti-bot handling	None	None	Built-in (stealth + auto-solve)
Challenge solving	None	None	Cloudflare + CAPTCHA autonomous
Content reading	Raw DOM	Snapshots	Smart multi-source extraction
Intent-based actions	Selectors only	Refs + intents	Natural language intents
Access state machine	No	No	Yes (READABLE, CHALLENGE, etc.)
MCP integration	No	No	Yes (9 MCP tools)

Quick Start

As MCP Server (for Claude, Cursor, etc.)

# Install
cd AgenticBrowser
npm install
npx playwright install chromium

# Run as MCP server
npm run mcp

Add to your MCP config:

{
  "mcpServers": {
    "agentic-browser": {
      "command": "npx",
      "args": ["tsx", "/path/to/AgenticBrowser/src/server.ts"]
    }
  }
}

As CLI

# Open a URL (auto-handles Cloudflare/CAPTCHA)
npx tsx src/cli.ts open https://example.com

# Read page content as markdown
npx tsx src/cli.ts read

# Click by intent
npx tsx src/cli.ts act click "the login button"

# Extract structured data
npx tsx src/cli.ts extract '{"products": [{"name": "", "price": ""}]}'

As SDK

import { openUrl, readContent, actOnPage } from 'agentic-browser';

// Open and auto-handle challenges
const page = await openUrl('https://example.com');

// Read clean markdown content
const content = await readContent({ format: 'markdown' });

// Click by natural language intent
await actOnPage({ action: 'click', intent: 'the pricing tab' });

MCP Tools

Tool	Description
browser_open	Open URL, auto-handle challenges, return state
browser_observe	Page summary + interactive elements
browser_read	Clean content extraction (markdown/text/html)
browser_act	Intent-based interaction (click, type, scroll, etc.)
browser_extract	Schema-driven structured data extraction
browser_verify	Goal/condition verification
browser_recover	Try alternative access methods
browser_debug	Console, network, screenshot diagnostics
browser_navigate	Back, forward, reload, goto

Stealth Strategy

The browser passes anti-bot systems through 6 layers:

1. Real Chrome binary — Authentic TLS fingerprint, HTTP/2, rendering
2. Automation marker removal — navigator.webdriver, chrome.runtime, cdc_ props
3. Realistic fingerprint — Consistent canvas, WebGL, hardware properties
4. Human-like behavior — Bézier mouse curves, variable keystroke timing
5. Autonomous challenge solving — Cloudflare auto-wait, CAPTCHA checkbox click, audio STT
6. Session persistence — Cookies and profile data persist across sessions

Architecture

Agent Layer (MCP / CLI / SDK)
    ↓
Command Router (9 commands)
    ↓
Autonomous Solver (Cloudflare + CAPTCHA)
    ↓
Access State Machine (classifies every page)
    ↓
Smart Reading Engine (multi-source extraction)
    ↓
Stealth Browser Runtime (Playwright + real Chrome)

Requirements

• Node.js 20+
• Chrome or Chromium installed
• TypeScript 5+

License

MIT