browser-use 1.0.2 — browser-use 1.0.2

browser-use 1.0.2 mainly removes a dead end that stopped its MCP server from working on many machines. browser-use is an open-source library that lets an AI model operate a real web browser, and since the 1.0 rewrite the model does it by writing Python against Chrome DevTools Protocol (CDP), the interface Chrome exposes so other programs can drive it. Its MCP server, which lets coding assistants and other AI apps call browser-use as a tool, can attach to the Chrome you already use, but only when Chrome's remote debugging setting is switched on at chrome://inspect. In 1.0.0 and 1.0.1, when it was off, the server stopped at instructions for turning it on. The pull request behind the release reports that about 390 Mac and Windows devices, and 72 of 101 devices running Cursor on Linux, got only that error and never drove a browser.

In 1.0.2, according to the release notes, the server instead launches its own Chromium, the open-source browser Chrome is built on, and tells the agent how to switch to your Chrome: tick Allow remote debugging at chrome://inspect/#remote-debugging and restart the MCP server. The difference that matters is logins. The pull request notes that the fallback Chromium has none until the toggle is on and the server restarted, so a task that needs your signed-in accounts still requires the switch.

Two smaller fixes concern waiting and text encoding. CDP calls from the command-line tool and MCP server now wait 20 seconds, as the agent already did, instead of 5, which gives slow pages more time but means a truly hung call takes 20 seconds to fail. A page frozen by a JavaScript dialog such as an alert box now fails within 2 seconds with an error saying to handle the dialog. On Windows, open() in code run through the MCP server's browser_exec tool now reads and writes UTF-8.

The release also adds ways to send feedback to the Browser Use team: a browser-use feedback command, a browser_use.feedback function, a browser_feedback tool for MCP clients, and a report_issue call the agent itself can make. The notes say the text is scrubbed before it leaves, and a commit in the 1.0.1 to 1.0.2 changes says paths and secrets are scrubbed. Agent and MCP reports are capped at three per run or per server, and feedback attempted from code the model writes is blocked. Neither the notes nor the pull request say whether the agent's own reports can be switched off.