browser-use: Browser automation surpasses Big Tech agents with 87.4% on Odysseys benchmark

An open-source framework that enables LLMs to manipulate web browsers like humans. It performs all interactions within the browser—such as opening pages, clicking buttons, entering text, and filling forms—using only natural language instructions. Unlike traditional scraping methods, it does not rely on DOM structures or selectors, but instead understands visual elements and context to complete tasks.

It achieved an average score of 87.4% on its own benchmark targeting 100 real-world browser tasks and on the Odysseys leaderboard. It distinguishes itself by demonstrating higher performance than computer use agents from OpenAI, Anthropic, Google, and Microsoft. It integrates with various LLM providers such as OpenAI, Anthropic, and Google via a Python library, and also offers models specialized for browser automation.

One-off tasks can be handled by integrating with existing coding agents via the CLI, while repetitive automation tasks can be scheduled or run in parallel using the Python library. It supports over 1,000 integrations including Gmail, Slack, and Notion. The cloud version provides proxy rotation, CAPTCHA solving, and stealth browser capabilities. It is distributed under the MIT license and also supports running local models.