Browser-use is an open-source Python library that gives a language model control of a real browser instead of an API.
What it is
A developer points Browser-use at Claude, GPT, Gemini, or a local model, and the library translates the model's decisions into clicks, typed input, and page reads. Under the hood it drives Chromium through Playwright, the same automation library Microsoft ships for cross-browser testing. That choice makes Browser-use fast to build on, but it also means every session inherits Playwright's own automation fingerprints unless something patches them first.
The project reports an 89.1% success rate on the WebVoyager benchmark, a suite of open-ended browsing tasks, which is why it has become the reference implementation a lot of other "AI browser agent" tools build on or copy from.
How it gets used to scrape
Nobody buys Browser-use to fill out one form. Teams wire it into agents that log in, navigate multi-step flows, extract data across paginated pages, and repeat the same session shape thousands of times an hour. The model reasons about what to click next from the rendered page, so the traffic looks like a real user clicking through a site, just faster and on a schedule no person keeps.
Why it's hard to catch with old checks
Because it runs on stock Playwright, a Browser-use session carries navigator.webdriver, the CDP-driven Runtime domain, and the other page-load tells that a decade of bot-detection blog posts already cover. That makes it catchable at the fingerprint layer, right up until someone bolts a stealth patch or a patched-Chromium build onto it, which is common in production deployments. The signal that survives that swap isn't the fingerprint. It's the shape of the session: how consistently the agent re-runs the same task, and how differently a model's decisions time out compared to a person reading the page.
Centinel detects it
In our September 2026 benchmark, Centinel identified every Browser-use session we ran against it, whether or not it carried a stealth patch. Centinel scores the connection, browser, and behavioral signals together across the session rather than trusting any single page-load check, which is what still catches an agent even after the obvious fingerprints are patched away.
Key takeaways
- Browser-use is an open-source framework that lets an LLM drive Chromium through Playwright, and it's become the default base for a lot of "AI browser agent" tooling.
- Its traffic inherits Playwright's fingerprints unless combined with a separate stealth or patched-Chromium layer.
- Behavioral consistency across a session, not a single fingerprint check, is what still identifies it once those patches are applied.