Go to a URL (or search Google with plain words), wait for the load, and return the page as a compact outline: headings plus interactive elements as [ref] role "name". Use the refs with browser_act.

Read the current page. view: page (outline with refs, default), text (readable content), links, forms, tables, meta. Use filter to narrow, offset to page.

Act on an element by ref with real input, wait for the page to settle, and get back only what changed. action: click, fill (text), press (key), check, uncheck, select (value, or values for a multiple list), hover, scroll (to: top|bottom, or ref), type (text at focus), upload (paths). Add submit:true to fill to press Enter. If the reply is approval.required or handoff.required, call browser_wait_human.

Run several browser_act steps in order in one call (e.g. fill a whole form and click Save). Stops at the first failure. Returns the final change set.

Wait until text appears, a CSS selector is visible, a selector is gone, or the URL contains a string. With no condition, waits for the page to finish loading and settle.

Manage your tabs. op: list, new (url), switch (tabId), close (tabId), back, forward, reload.

Screenshot the page (or one element by ref) and return the image. Costs many tokens: prefer browser_read unless you need to see layout, images or a canvas.

Wait for the human operator: pass approvalId after approval.required, or nothing after handoff.required (CAPTCHA, login). Never try to solve a CAPTCHA yourself.

How to use this browser, written for you. Call it once before your first browser action. topic=start: the working loop, rules that save turns, errors, accounts (short; read this first). topic=commands: find commands: pass query (words, e.g. "selector click") for the matching ones, or group for one section (open, read, act, batch, accounts, person, browser; all = every command, about 25,000 tokens). topic=gomini: GoMini, the optional local model that takes routine steps, and how to phrase subgoals for it.

Use the accounts the person linked for you in the Account Center. op=accounts lists them (linked or not); op=apis lists the APIs you can call; op=call sends one API request as the account (connector, method, path, body) without you ever holding a token; op=sign_in has the browser type the saved sign-in into the current page's login form. If an account is not linked you get policy.denied: ask the person to link it, do not work around it.

Call any other AgentGo protocol method directly (cookies, downloads, PDF, dialogs, network log, workflow.collect ...). browser_call {method:"browser.capabilities"} lists them.