Browser Use is an open-source browser agent that operates a web browser the way a person does: it opens pages, clicks, types, fills in forms, and works through multi-step tasks such as finding a free slot and booking an appointment from a natural-language task description. The open-source agent is available for Python and TypeScript.
It comes in three forms. A fully hosted cloud API runs both the agent and the browser and adds stealth browsers, residential proxies, CAPTCHA handling, profiles, and recordings. A CLI, installed into an existing coding agent such as Claude Code, Codex, Hermes, OpenClaw, or Pi through a setup prompt, gives that agent control of a browser. The Python library runs the Browser Use agent inside your own application with custom tools, structured output, and your choice of model. The CLI and library each connect to a local browser, a cloud browser, or an existing remote browser over the Chrome DevTools Protocol (CDP).
The library works with the project's own BU2 model, tuned for browser automation, or with providers directly through wrappers for OpenAI, Anthropic, and Google, and with local models through Ollama. A gateway also accepts provider-prefixed model IDs for Claude, GPT, and Gemini. The agent supplies its own system prompt, which you can extend or replace, and custom tools are registered as plain functions.
A Claude toolset integration lets Claude's browser tool use Browser Use as the driver, implementing all 31 browser actions, with a Bash tool for processing data and writing files. The project also publishes a hard browser-task benchmark that compares agent success rate and cost per task across models.
Features
- Natural-language tasks: the agent receives a task description and drives the browser to complete it without a scripted click path
- Hosted cloud: an API that runs the agent and browser, with stealth browsers, proxies, profiles, and recordings
- Coding-agent CLI: gives Claude Code, Codex, Hermes, OpenClaw, Pi, Cursor, and other agents control of a browser
- Python library: embeds the open-source agent in your own code with custom tools and structured output
- Browser choice: a local browser, a reused Chrome profile, a cloud browser, or a remote browser over CDP
- Model choice: the BU2 browser-automation model, OpenAI, Anthropic, and Google wrappers, or local models through Ollama
- Custom tools: register any function as an action the agent can call
- Claude toolset: drives Claude's browser tool with all 31 browser actions plus a Bash tool
- Benchmark: a published set of hard browser tasks comparing success rate and cost per task by model
