Browser automation where LLMs drive the mouse

What it does

Browser-use is a Python library that hands an LLM the keys to a real Chromium session via Playwright. You describe a task in plain English—apply for a job, buy groceries, scrape a site—and the library coordinates the model, the DOM, and the clicks until the task is done or the budget runs out. It supports OpenAI, Anthropic, Google, local Ollama models, and the project’s own hosted ChatBrowserUse endpoint.

The interesting bit

The project doesn’t hide the fact that it’s also a sales funnel for Browser Use Cloud. The open-source library is MIT-licensed and extensible with custom tools, but the README steadily nudges you toward the hosted version for stealth browsers, CAPTCHA avoidance, and production scaling. That tension makes the repo unusually honest about its limits.

Key highlights

• Ships with a persistent CLI for iterative browser control—navigate, click elements by index, type text, and capture screenshots without writing scripts.
• Exposes an open benchmark suite (browser-use/benchmark) covering 100 real-world browser tasks.
• Lets you register custom Python tools with a @tools.action decorator to extend the agent’s capabilities.
• Supports real browser profiles so the agent can reuse existing cookies and logins instead of starting cold.
• Includes a Claude Code skill and project templates for quick scaffolding.

Caveats

• The FAQ explicitly warns that production workloads with many parallel agents are tricky due to Chrome’s memory appetite, and pushes users toward the paid Cloud API.
• CAPTCHA handling and advanced stealth fingerprinting are gated behind the Cloud offering; the open-source agent alone is not expected to evade serious bot detection.
• The hosted ChatBrowserUse model claims 3–5× faster task completion than other LLMs, but that metric comes from the project’s own benchmark.

Verdict

Worth a look if you need an LLM to interact with websites that lack APIs—especially for prototyping, form filling, or personal automation. Pass if you need undetectable, high-scale scraping; you’ll end up on the Cloud pricing page before you finish the README.