browser-use 1.0 is the first stable release of browser-use, the open-source browser agent. It shipped on November 10, 2026, and the official blog post on November 20 explained the thinking behind the rewrite. The biggest change is not a new feature but a subtraction: the 24 browser actions of 0.x have been merged into a single python tool, letting the model write Python directly in a persistent REPL and drive the browser through the Chrome DevTools Protocol. According to the official post, the rewrite deleted about 137,000 lines of code, shrinking the library from 178 Python files to 43. The official repository is on GitHub, under the browser-use organization, with the project name browser-use.
What really changed when 24 tools became one
In 0.x the model could only pick one action per step from a menu, and it got stuck whenever the menu had no answer. In 1.0 the model writes code directly: if a page hangs it can stop the load, block the stuck script, and read the page's own API to fetch data — in one example from the official post, the agent retrieved 4,720 rows of data in three and a half minutes this way. The trade-off is that debugging changes: when something goes wrong you inspect the code the model wrote, not the parameters of one action.
Three changes to check before upgrading
First, old code will most likely not run: removed 0.x imports raise an ImportError that points to the migration guide, and the official advice is to read MIGRATING.md before upgrading. Second, the defaults now point at the cloud: with no model specified it uses bu-fast on the gateway, and with no browser specified it uses a Browser Use Cloud browser — local Chromium must be requested explicitly with Browser('chromium'), and local network addresses only work locally. Third, the step limit has been tightened from 500 to 100 steps, while API keys have been consolidated into a single BROWSER_USE_API_KEY.
Deployment cost, real pitfalls, and who it suits
Running it locally needs only a Python environment and a browser; the money goes mainly to model calls and the cloud browser, and cloud browser and proxy fees are not included in result.cost_usd, so they must be tracked separately. There are four real pitfalls: login walls and captchas can still stop the agent — in the official test the dedicated captcha tool passed 6 out of 6 times while a prompt alone failed 3 out of 3; long tasks can hit the 100-step limit; all dependencies are pinned to exact versions, which can conflict with an existing project; and cloud browsers do not yet support viewport emulation. It suits developers and small teams who want to automate web data collection, form filling, and repetitive browser work in bulk. It is not for ordinary users who just want to click through a few pages, nor for scenarios sensitive to data leaving the local network. The alternatives are script-it-yourself frameworks such as Playwright and Selenium.