browser-use 1.0.3 — browser-use 1.0.3

Browser Use 1.0.3 is a reliability patch, and most fixes in it target a failure the company counted on machines running 1.0.2. Browser Use is an open-source Python library that lets an AI model operate a real web browser, from Python code, from its command-line tool, or through its `--cli-mcp` server, which coding assistants and other AI apps call as a tool. The pull request behind the release lists, for most changes, how many devices hit the problem in 1.0.2 error data collected over about eleven hours on November 11. Those numbers are Browser Use's own telemetry, so they show where the vendor saw failures rather than an independent measure of how common they are. The largest, 373 devices, was the MCP server exiting with an error when the calling app closed the connection before the reply arrived, a bug the pull request says was also in 1.0.1; in 1.0.3 the server exits cleanly.

The change command-line users are most likely to notice extends what 1.0.2 did for the MCP server. The command-line tool failed when no browser had remote debugging switched on (the Chrome setting that lets other programs drive it) or when the machine had no display, and the pull request counts 171 devices that never got a browser, 58 of them headless Linux machines. In 1.0.3 the tool starts its own Chromium on a fresh temporary profile instead, headless when there is no display, keeps it running across calls, and closes it on `browser-use --reload` (release notes). The trade is the same one 1.0.2 made: this is not your Chrome, so it has none of your signed-in accounts, and the tool prints a notice saying so. The pull request also says the fallback was not run on Windows, macOS or a desktop with a display, and that on a desktop the Chromium opens as a visible window that stays until the reload.

Two smaller fixes shorten or remove other dead ends when connecting to a browser. If the BU_CDP_URL setting points at an address where nothing is listening, browser-use now fails in 3 seconds with one message naming the value and both ways out, instead of waiting 30 seconds; the pull request counts 108 devices that hit this. Chrome running in a separate container, or Cursor on Linux, is now accepted when the browser answering on the port Chrome recorded is the same one that wrote that record, which the pull request says about 24 Linux and Cursor devices were refused before. It also says this fix likely does not help branded Chrome 147 and later, only Chromium and browsers built on it.

The fix with money attached concerns cancelled runs on Browser Use's cloud browsers. When such a run stops, browser-use copies the files the cloud browser downloaded back to the local machine, and each listing and copy pass could hold the stop, and with it the cloud billing, for up to 300 seconds. A cancelled run, whether through task.cancel(), Ctrl-C or a timeout, now gets one shared 10-second budget for all of those passes. The pull request states the cost: a download that cannot be listed and copied within 10 seconds is lost, or, for a run set to keep its browser alive, goes to the next run's workspace, and either case is logged. It also says each fix was checked by reproducing the failure before and after with throwaway tests that were not committed, so the evidence that these failures are gone is the vendor's own. For anyone who hit one of them on 1.0.2, the upgrade is narrow, and the stated rollback is pinning 1.0.2 (all changes).