Authority AI Editorial
Browser Use: agents that drive a real browser
Browser Use is an open-source library that gives an AI agent control of a real web browser. Instead of scraping a page or calling an API, the agent looks at the page, decides what to click or type next, and works through a task the way a person would: filling in a form, clicking through a multi-step checkout, or finding a specific piece of information buried behind a few navigation steps. It's aimed at people building agents that need to operate on websites that were never built with an API in mind, which covers a large share of the internet. For someone already running coding agents, it's the tool to reach for when the task is "go do something on this website," not "call this endpoint."
What it does
You give a Browser Use agent a task as a plain-language instruction, point it at a model, and it drives an actual browser session step by step: reading the page, choosing an action, taking it, and checking the result before deciding what to do next. The project's README describes handling booking flows, CAPTCHAs, form filling, and other multi-step web tasks this way. It's released under the MIT license, is written primarily in Python, and also ships a TypeScript variant for projects that live in that ecosystem.
How you use it
The quickest path is the open-source Python library: pip install browser-use, then uvx browser-use install to pull down a Chromium build, on Python 3.12 in a virtual environment. From there a minimal script imports Agent and a chat model, wraps a one-line task description, and calls agent.run(). The docs show a working example using their own ChatBrowserUse model, and also support pointing the agent at Gemini, OpenAI, or Anthropic Claude with your own API key, so you aren't locked into one model provider. For anything that needs a logged-in session held across runs, the docs point to a @sandbox decorator backed by a cloud profile ID that manages the agent, the browser, persistence, authentication, and cookies as one unit rather than something you wire up by hand.
What to know before adopting it
The open-source library is free; you only pay for whichever model's inference you route through it. If you'd rather not run browser infrastructure yourself, Browser Use also offers a fully hosted cloud API, priced on the docs at $0.02 per browser-hour, with stealth and CAPTCHA-handling built in. The project also has a CLI path for connecting an existing coding agent, such as Claude Code, to a local or cloud browser rather than writing a new Browser Use script from scratch. As of 2026-09-27 the repository shows 117k stars and 12.8k forks, with more than 10,299 commits, and it ships its own benchmark suite for comparing model performance on browser tasks. The project recommends its own model but works with the major providers, so you can start with what you already have a key for.
Where it sits next to similar tools
Browser Use solves a narrower problem than a general agent framework: it doesn't orchestrate multi-agent systems or long-running workflows on its own, it drives a browser. That makes it a good fit alongside a broader framework, where Browser Use handles the "operate this website" step and something else handles the surrounding logic. If your target already has a clean API, a browser-driving agent is the wrong tool: you'd be adding screen-reading overhead to a job an HTTP request already does. Where a full agent needs ready-made methods rather than ones written from scratch, a packaged set of skills can fill that gap; the AI Ultimate Skill Collection has 3,029 of them in 12 packs, written procedures your agent follows step by step, which is a different kind of building block than a browser driver but saves the same work of writing every method by hand.