Browser-Use

Browser-Use is a popular open-source Python library that enables AI agents to autonomously control web browsers using natural language instructions. Built on top of Playwright and integrated with LangChain, it allows LLMs like GPT-4o and Claude to navigate websites, fill forms, extract data, and perform complex multi-step web tasks. With over 50,000 GitHub stars, it has become the leading framework for agent-driven browser automation.

Architecture

Browser-Use follows a modular, agent-based architecture with three core components:

Agent — The central orchestrator that takes a natural language task, an LLM, and a browser session. It autonomously reasons about page content and decides actions (click, type, scroll, navigate).
BrowserSession — Manages browser connections via Chrome DevTools Protocol (CDP). Supports local Playwright browsers or cloud-hosted browsers via Browserless WebSocket endpoints. Configurable via BrowserProfile for headless mode, viewport size, and user-agent.
LLM Integration — Uses LangChain-compatible chat models (ChatOpenAI, ChatAnthropic) for decision-making. The LLM interprets DOM content, screenshots, and page state to determine the next action.

The agent loop works as follows: observe the page state (DOM + optional screenshot) → send to LLM → receive action → execute via Playwright → repeat until task complete.

How It Works with Playwright

Browser-Use relies on Playwright as its browser automation engine. Rather than requiring developers to write Playwright scripts, the library abstracts browser control behind the Agent interface:

Playwright launches Chromium, Firefox, or WebKit browsers
The agent connects via CDP (Chrome DevTools Protocol) for real-time control
Actions like clicking, typing, scrolling, and navigation are executed through Playwright's async API
Screenshots are captured for vision-capable LLMs to analyze
DOM extraction provides structured page content for text-based reasoning

For cloud deployments, Browser-Use connects to Browserless or similar services via WebSocket CDP URLs, avoiding the need for local browser installations.

Key Features

Multi-Tab Browsing — Agents can open and manage multiple browser tabs simultaneously for tasks like comparison shopping
Vision Capabilities — GPT-4o and other vision models analyze screenshots for visual reasoning alongside DOM text
DOM Extraction — Full DOM tree parsing with intelligent element selection for LLM consumption
Custom Actions — Define custom action handlers for domain-specific interactions
Structured Output — Pydantic schema support for typed, validated extraction results
Parallel Agents — Run multiple agents concurrently for cross-site tasks
Async/Streaming — Real-time step-by-step visibility into agent actions

Integration with LangChain and OpenAI

Browser-Use is designed as a LangChain-native tool:

Uses langchain_openai.ChatOpenAI or langchain_anthropic.ChatAnthropic as the reasoning engine
Compatible with any LangChain-compatible LLM provider
Agents can be embedded into larger LangChain chains and workflows
Supports OpenAI function calling for structured tool use

Code Example

import asyncio
from dotenv import load_dotenv
from langchain_openai import ChatOpenAI
from browser_use import Agent, BrowserSession, BrowserProfile
 
load_dotenv()
 
async def main():
    # Configure browser session
    session = BrowserSession(
        browser_profile=BrowserProfile(headless=True)
    )
 
    # Create agent with GPT-4o
    agent = Agent(
        task="Go to Hacker News, find the top post, and return its title and URL.",
        llm=ChatOpenAI(model="gpt-4o"),
        browser=session,
    )
 
    # Run the agent
    result = await agent.run()
    print(f"Result: {result}")
 
asyncio.run(main())

Architecture Diagram

graph TD A["User Task (natural language)"] --> B["Agent (reasoning loop)"] B --> C["LLM (GPT-4o / Claude)"] B --> D["Browser Session (Playwright CDP)"] C -->|interprets page state| B D --> E["Browser (Chromium / Firefox)"] E -->|DOM + screenshots| B

Table of Contents