bopbop

Tools

What your agent can call. You do not call these yourself: ask the agent in plain words.

Tool Use
navigate Open a URL (or back, forward, reload). Returns the snapshot; snapshot: interactive|none for less.
act Batched steps; returns step results, events (navigation, dialogs, new tabs, errors) and what changed. Refs or locators: text=Save, css=.bar button, e12 >> text=Save; alternatives css=textarea, text=Ask (first match wins); an ambiguous text= gives a warning.
snapshot Page structure with refs. search, ref, viewport, mode=interactive, scope=page (default: the open dialog).
read Page text. format=markdown: the main content as Markdown.
extract Tables as rows keyed by column, label/value tables and <dl> as fields, as JSON.
wait Until text appears or goes, a ref is visible, or the URL matches.
screenshot JPEG of the viewport, a ref/locator, a clip rect (+padding) or the full page; marks draws refs. 1280 px wide at most, scale: full for 1:1. Returns the captured area in CSS px.
eval JavaScript for values that the other tools do not show. Awaits a returned Promise (timeout). One line if the page changed; snapshot: diff|full shows the change.
console Console messages and page errors; identical messages show once with ×N.
network Requests with time and size; id: phases, headers, bodies; cookies; har export.
emulate Device presets, viewport, dark mode, reduced motion, network and CPU throttling, reset; window=true resizes the agent window.
perf Web Vitals with ratings, long tasks, slowest and largest resources, heap and DOM size.
tabs List, open, select (one line), close, duplicate (keeps sessionStorage). Extension mode: label names the tab group Agent:<1-5 letters, digits or emoji> (default: project initials + slot, e.g. Agent:CT2).

act steps: click, type (+submit), press, hover, select (native or custom combobox, verified), check/uncheck, upload, drag (+position), click_xy, scroll (+until), wait, wait_for, dialog, navigate. Any ref field also accepts text=Label.