browser-use
A Python library that gives an agent a browser and a task in plain words: go here, find this, fill that in. A hundred thousand stars, MIT.
Half of the routine on the internet is not thinking but repetition: collect prices from ten sites, pull a report from a dashboard with no API, fill the same form forty times. browser-use takes that part: you describe the task in words, the agent opens a browser and works through the steps itself.
How it differs from Playwright MCP, its neighbour in this catalogue: there the browser is handed to the model as tools inside your conversation with Claude Code. Here it is a Python library you build your own script or service from. The first is convenient while you sit and watch; the second is what you want when the job has to run on a schedule, without you.
What it costs. An agent in a browser is slower and dearer than a direct API call: every step is a screenshot or a page tree sent to the model. If the service has an API, use the API. browser-use earns its place exactly where there is none, or where it is closed.
And the same warning as for any agent with outward access: it walks the network as you. A separate browser profile, separate accounts, no sessions carrying payment details — not until you have watched what steps it actually takes.
Comments