BrowserSkill review: setup, browser access and risks

BrowserSkill is Tencent's open source tool for connecting a coding assistant to a browser session where the user is already logged in. The video explains its local command-line program, background daemon and browser extension. The speaker names Claude Code, Cursor and Codex as supported agents and says other agents can drive it through shell commands.

The setup includes a diagnostic check, but the user must install the extension manually. Agents borrow tabs and work in a separate visible window while the user continues browsing. The speaker describes access to authenticated dashboards and company wikis, with a human handoff when a CAPTCHA appears. This is browser automation on the user's machine; the video does not establish that the driving model runs locally.

Much of the review concerns permissions. The speaker argues that the borrow-and-return rule depends on agent compliance rather than a hard tab boundary, and explains how malicious page text could redirect an agent through prompt injection.

The review discusses version 0.2.0 and reports six stable core test cases covering a tool with 28 browser operations. It also describes connection loss after form submission and a Windows session-start hang despite passing diagnostics. The speaker says the form issue is marked for 0.2.1, rather than presenting that fix as shipped. The closing advice bases delegation on the consequences of a wrong click, especially when an action sends a message or cannot be undone.