🌐 Browser-Use

Open-Source Web Automation with LLMs

Let LLMs interact with websites through a simple interface.

Short Example

pip install browser-use

from langchain_openai import ChatOpenAI
from browser_use import Agent

agent = Agent(
    task="Go to hackernews on show hn and give me top 10 post titles, their points and hours. Calculate for each the ratio of points per hour.",
    llm=ChatOpenAI(model="gpt-4o"),
)

# ... inside an async function
await agent.run()

Demo

Prompt: Go to hackernews on show hn and give me top 10 post titles, their points and hours. Calculate for each the ratio of points per hour. (1x speed)

Prompt: Search the top 3 AI companies 2024 and find what out what concrete hardware each is using for their model. (1x speed)

Prompt: Go to kayak.com and find a one-way flight from Zürich to San Francisco on 12 January 2025. (2.5x speed)

Prompt: Opening new tabs and searching for images for these people: Albert Einstein, Oprah Winfrey, Steve Jobs. (2.5x speed)

Local Setup

1. Create a virtual environment and install dependencies:

# I recommend using uv
pip install .

2. Add your API keys to the .env file:

cp .env.example .env

E.g. for OpenAI:

OPENAI_API_KEY=

You can use any LLM model supported by LangChain by adding the appropriate environment variables. See langchain models for available options.

Features

• Universal LLM Support - Works with any Language Model
• Interactive Element Detection - Automatically finds interactive elements
• Multi-Tab Management - Seamless handling of browser tabs
• XPath Extraction for scraping functions - No more manual DevTools inspection
• Vision Model Support - Process visual page information
• Customizable Actions - Add your own browser interactions (e.g. add data to database which the LLM can use)
• Handles dynamic content - dont worry about cookies or changing content
• Chain-of-thought prompting with memory - Solve long-term tasks
• Self-correcting - If the LLM makes a mistake, the agent will self-correct its actions

Advanced Examples

Chain of Agents

You can persist the browser across multiple agents and chain them together.

from asyncio import run
from browser_use import Agent, Controller
from dotenv import load_dotenv
from langchain_anthropic import ChatAnthropic
load_dotenv()

# Persist browser state across agents
controller = Controller()

# Initialize browser agent
agent1 = Agent(
    task="Open 3 VCs websites in the New York area.",
    llm=ChatAnthropic(model="claude-3-5-sonnet-20240620", timeout=25, stop=None),
    controller=controller)
agent2 = Agent(
    task="Give me the names of the founders of the companies in all tabs.",
    llm=ChatAnthropic(model="claude-3-5-sonnet-20240620", timeout=25, stop=None),
    controller=controller)

run(agent1.run())
founders, history = run(agent2.run())

print(founders)

You can use the history to run the agents again deterministically.

Command Line Usage

Run examples directly from the command line (clone the repo first):

python examples/try.py "Your query here" --provider [openai|anthropic]

Anthropic

You need to add ANTHROPIC_API_KEY to your environment variables. Example usage:

python examples/try.py "Search the top 3 AI companies 2024 and find out in 3 new tabs what hardware each is using for their models" --provider anthropic

OpenAI

You need to add OPENAI_API_KEY to your environment variables. Example usage:

python examples/try.py "Go to hackernews on show hn and give me top 10 post titles, their points and hours. Calculate for each the ratio of points per hour. " --provider anthropic

🤖 Supported Models

All LangChain chat models are supported. Tested with:

• GPT-4o
• GPT-4o Mini
• Claude 3.5 Sonnet
• LLama 3.1 405B

Limitations

• When extracting page content, the message length increases and the LLM gets slower.
• Currently one agent costs about 0.01$
• Sometimes it tries to repeat the same task over and over again.
• Some elements might not be extracted which you want to interact with.
• What should we focus on the most?
  • Robustness
  • Speed
  • Cost reduction

Roadmap

• Save agent actions and execute them deterministically
• Pydantic forced output
• Third party SERP API for faster Google Search results
• Multi-step action execution to increase speed
• Test on mind2web dataset
• Add more browser actions

Contributing

Contributions are welcome! Feel free to open issues for bugs or feature requests.

Feel free to join the Discord for discussions and support.

---

Star ⭐ this repo if you find it useful!
Made with ❤️ by the Browser-Use team