Skip to content

Browser-Use Automation — Setup Guide

Source: browser-use/browser-use — 97,000+ ★ | Python Category: Automation / Browser-Use Quality Tier: 🟢 Production (97K+ GitHub stars, active development)

Browser-use replaces Playwright/Selenium for sites defended by anti-bot systems. Instead of writing brittle selectors, you describe the task in natural language and a vision-capable LLM drives the browser like a human. It handles CAPTCHAs, JS-heavy rendering, and anti-bot detection natively.

How It Works

  1. You describe a goal ("log in to this site, navigate to the pricing page, extract the plans")
  2. A vision-capable model (Claude Sonnet, GPT-4 class) screenshots the page
  3. The model decides the next action: click, type, scroll, navigate
  4. Actions execute through a real browser until the task completes

This closed loop is why browser-use succeeds where selector-based automation fails: it adapts to dynamic DOMs, obfuscated class names, and layout shifts.

Installation

python3 -m venv venv-browser
source venv-browser/bin/activate
pip install browser-use
python3 -m playwright install chromium   # browser-use drives Playwright under the hood

Quick Start

import asyncio, os
from browser_use import Agent, ChatAnthropic

async def main():
    llm = ChatAnthropic(
        model="claude-sonnet-4-20250514",
        api_key=os.getenv("ANTHROPIC_API_KEY"),
    )
    agent = Agent(
        task="Go to linkedin.com/login, log in, navigate to jobs",
        llm=llm,
    )
    result = await agent.run()
    print(result)

asyncio.run(main())

Requirements: - Python 3.11+ - A vision-capable LLM API key (browser-use is LLM-agnostic: Anthropic, OpenAI, Gemini, or local Ollama models) - ~5 GB free disk for browser binaries

When to Use Browser-Use

Use browser-use for Stick with plain Playwright for
Sites with Cloudflare / anti-bot walls Simple form filling on undefended sites
Dynamic SPAs with obfuscated selectors API-authenticated platforms
CAPTCHA-protected flows Bulk scraping (LLM calls add cost)
Multi-step logins with security checkpoints Speed-critical extraction

Cost note: Every action costs an LLM call. For high-volume scraping, plain Playwright or direct APIs are far cheaper. Browser-use shines on flows where automation gets blocked, not where it merely works.

Headless Limitations

Anti-bot systems fingerprint headless browsers. Browser-use works best when the runtime environment looks like a real user:

  • Residential or mobile IPs pass Cloudflare checks where datacenter IPs get challenged
  • Headful mode with persistent profiles survives platform login flows (cookies + crypto-bound session keys persist together)
  • On headless servers, pair with Xvfb for a virtual display
Xvfb :99 -screen 0 1920x1080x24 -ac &
export DISPLAY=:99

Verification

python3 -c "import browser_use; print('browser-use ready')"

A successful smoke test: run the Quick Start against a public site and confirm the agent extracts the expected data.

Notes

  • Production-grade: 97K+ GitHub stars, used across the AI automation ecosystem
  • Vision-first: No CSS selectors needed — the model sees the page like a human
  • Best for: Anti-bot-defended platforms, dynamic SPAs, CAPTCHA flows
  • Complementary: Use alongside plain Playwright — browser-use for vision tasks, Playwright for cheap API-level control
  • Related: See Playwright Social Media Automation and Midscene

Setup guide by CorpusIQ. Source: browser-use/browser-use.