Browser Use

An open-source project by the Browser Use team (browser-use/browser-use). It’s the engine that powers the browser-use.com cloud.

What it does

To have an AI agent “handle something on the web for me,” you need a control layer that opens the browser, clicks, types, and reads results back. Handling each site’s JavaScript rendering, login session, and dynamic markup yourself is a heavy operational load.

Browser Use bundles that control layer into one framework. It gives agents direct, dependable control surfaces over the browser so they can autonomously carry out a task described in natural language. From the README: “Make websites accessible for AI agents.”

What you can automate

The use cases highlighted by the docs and browser-use.com:

• Form-filling, shopping, booking — multi-step web flows the agent completes end to end
• Research — gathering and organizing information across multiple sites
• Data extraction — pulling structured data out of pages
• Testing and monitoring — repeatedly checking how a web service behaves

The flow is: you describe a goal in natural language, and the agent reads the page and plans the clicks, inputs, and navigation on its own.

Key features

• Multi-LLM support

  Choose Anthropic, OpenAI, or Google models, or the purpose-built ChatBrowserUse() model for browser automation. Swap models to run or benchmark the same task.

• Real browser profile support

  Use an existing browser profile to work while staying logged in, keeping sessions across authenticated sites.

• CLI and custom tool integration

  Drive the agent directly from the CLI, or register custom tools to extend its behavior.

• Headless, parallel, and cloud execution

  Run many browsers headless for unattended automation, or execute on remote browsers via the browser-use.com cloud.

• Claude Code skill

  Register the skill with browser-use skill and control the browser through the browser-use harness from inside Claude Code.

How it differs from Claude in Chrome

Claude Code already has Claude in Chrome for driving your browser, so the two look overlapping — but they aim at different things.

Claude in Chrome is the supervised tool that “lets Claude drive the browser in front of you”; Browser Use is the automation framework that “runs code-defined browser agents freely across models, unattended and at scale.” Reach for Claude in Chrome for ad-hoc, supervised work on sites you’re already logged into; reach for Browser Use when you need unattended, high-volume, model-agnostic runs.

Usage

Install — add the package with uv or pip.

uv add browser-use
# or
pip install browser-use

Pick a model — configure the agent with its own model or any LLM you prefer.

from browser_use import ChatBrowserUse
llm = ChatBrowserUse(model='anthropic/claude-sonnet-4-6')

Use it from Claude Code — register the skill to drive the browser inside Claude Code. Python 3.12 is recommended.

browser-use skill

How it relates to browser-use.com

This repo (the open-source core) is MIT and works fully locally, but it’s also the engine behind browser-use.com, the commercial cloud run by the same team. The cloud is a paid service that layers on the stealth browsers, proxies, CAPTCHA solving, and remote execution (Browser Use Box) that are hard to self-host. The core runs entirely locally without the cloud.

Notes

• Python 3.11+ required — Python 3.12 is recommended for the CLI skill setup.
• LLM API key required — you need an API key for the model you run the agent with (including its own ChatBrowserUse).
• Cloud is optional — the core is MIT open source and works fully locally. Stealth browsers, proxies, and CAPTCHA solving are commercial features of the browser-use.com cloud.
• MIT License