Let LLMs interact with websites through a simple interface.

pip install browser-use

fromlangchain_openaiimportChatOpenAIfrombrowser_useimportAgentagent=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 functionawaitagent.run()

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)

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.

• 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

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

fromasyncioimportrunfrombrowser_useimportAgent, Controllerfromdotenvimportload_dotenvfromlangchain_anthropicimportChatAnthropicload_dotenv() # Persist browser state across agentscontroller=Controller() # Initialize browser agentagent1=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.

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

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

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

All LangChain chat models are supported. Tested with:

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

• 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

• 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

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

Feel free to join the Discord for discussions and support.