Operate a website autonomously in the user's real, logged-in Chrome: open url, then pursue goals written as plain sentences, deciding every click, selection and keystroke without asking. Right for multi-step work on one site: filling a form, walking to a page that is only reachable by clicking, searching inside a site, reading something behind a session the user already has. Wrong for reading one page (use browser_read), for public pages that need no login (use a web fetch or search tool), and for logging in, paying, ordering or submitting anything on a banking or payment site. Never put credentials into goals: the goals and every typed value come back in the result. The run stays on the registrable domain of url and stops when a step would leave it; widen that with allow_domains. Bounded by max_actions (default 25, values above 60 are rejected) and time_budget_s (default 120 s, values above 900 s are rejected); the call blocks until the run ends, and only one run can be in flight at a time, including any run still finishing after its budget ran out. Set dry_run to get the first planned step without executing anything. The underlying agent cannot see into iframes or shadow DOM, cannot do file uploads, cannot follow pop-up tabs, and only observes the visible viewport, so it scrolls but never reads what is not rendered. The result reports the status, the final URL, every executed step and the reason it stopped.

Report whether a browser run can work at all, before spending time on one. Says which API keys are present and where they came from, which text model would type into fields, whether the browser-harness daemon is running and whether a Chrome is connected to it, and which operations are therefore blocked. Read-only and free: it opens no browser, loads no page, calls no model and costs nothing. Call it when browser_task or browser_read fails, before the first run of a session, or when the user asks whether the browser agent is ready. It says nothing about any particular website.

Open one page in the user's real, logged-in Chrome and return its visible text plus the table of elements that can be operated. This tool clicks nothing, types nothing and selects nothing. Opening the page is itself a request, though: it opens a background tab and navigates to url with the user's cookies, so a URL that acts on GET will act, for example a logout, an unsubscribe, a confirmation link or a one-click action. Right for reading or checking a page, for seeing what a page offers before starting browser_task, and for content behind a login the user already has. Wrong when anything has to be clicked or filled in (use browser_task), and wasteful for public pages that need no login (use a web fetch or search tool). Redirects are followed but verified: landing on another registrable domain is reported, and none of that page's text, elements, title or address is returned; widen that with allow_domains. The underlying library observes at most 6000 characters of visible text and cuts a longer page before text_limit (default 4000) applies, so the answer reports how long the observed text was and warns when it reached that ceiling. Labels and field values are capped at 200 characters and option lists at 50 entries per element, and the answer says when it capped something. Same limits as the agent: no iframes, no shadow DOM, no file uploads, no pop-up tabs, and only the visible viewport is observed, so content further down the page can be missing. Only one run can be in flight at a time.