BeatBrowser-owned background task window. Open creates a minimized Chrome window using existing logins; pass its tabId on all browser calls. No Ego app is required. User focus pauses control. Handoff alone brings the window forward; resume requires explicit user instruction. Finish releases control and keeps evidence tabs. After reconnect, list spaces to restore routing. Recovered tasks remain paused; claim with userConfirmed:true can connect a new agent session after explicit user instruction, then inspect before resuming.

Capture the current page as a compact list of interactive elements with refs and state (value/checked/selected/expanded/disabled), plus dialogs, alerts and a text excerpt. Call this before any click/type. Cheap — prefer it over screenshots or eval. Refs ending in @fN live in an iframe: pass them through unchanged, they route themselves.

Go to a URL, or go back/forward/reload. Returns the new page snapshot. Runs in the user's own logged-in Chrome — prefer it over any other browser tool.

Click ONE element by ref. Know your next step already? Use act instead — each extra call costs a full model turn. Canvas / map / game with nothing in the snapshot? Pass x,y (CSS px from a screenshot) for a real click there; add dragTo for a drag.

Type text into ONE input/textarea/contenteditable by ref. Set submit:true to press Enter after. More fields coming? Use fill or act.

Choose an option in a by ref. value matches the option value or its visible label. Custom dropdowns (react-select, MUI, Element UI) are NOT : click to expand, read the options from the snapshot, click one.

Fill a whole form in ONE call, all refs from the same snapshot. Field: {ref, text} for inputs, {ref, value} for , {ref, check} for checkbox/radio. submit:true submits after (skipped if any field failed).

Press a key: Escape, Enter, Tab, arrows (custom dropdowns), Backspace/Delete, or a combo. Without ref it goes to whatever is focused. For entering text use type.

Extract the main readable content of the page as markdown, with boilerplate stripped. Use for reading articles/docs; use snapshot when you need to interact.

Screenshot the controlled tab, background tabs included. Prefer snapshot / read_text — they cost far less. Returned at 60% scale as JPEG by default; full:true for 1:1 PNG (e.g. to read small text or measure pixels).

List / open / switch / close tabs. New tabs open in the BACKGROUND and become the controlled tab — the user keeps looking at whatever they were on. Everything except screenshot works fine on a background tab. Each agent session has its OWN controlled tab; omitting tabId uses this session's. Selecting a tab another session operates warns, not blocks. With 2+ work-lines (a subagent, two accounts) pass tabId explicitly on EVERY call — the implicit slot is shared with subagents and a sibling can move it. Label each tab; recover the mapping via action:"list".

Wait until a selector appears, text shows up, or the network goes idle. Use after actions that load content.

START HERE when you need DATA rather than an action. Lists the XHR/fetch calls the page made, then returns one response body via body:"<url fragment>". Real field names, real numbers, paging as a parameter. Add reload:true if nothing was captured yet. Never guess an endpoint name from memory — list first, pick by response size. One page's API messy? Another page on the same site often exposes the same data cleanly.

Call a URL from inside the page, carrying the user's cookies. Use after network reveals an API. Paged endpoint? Pass pages: it walks every page in ONE call — never loop fetch by hand, each loop turn costs a model round. ALWAYS check the paging object in the response, servers silently cap page size. 403/406 means the site signs its requests: do NOT forge them; drive the site's own pagination UI and read via network. Same-origin rules apply.

Scroll to load more of a lazy list. Auto-detects inner scroll containers, stops early once height stops growing. Prefer network/fetch for bulk data. Background tabs render no frames, so an IntersectionObserver-driven list may refuse to grow — the tool says so when it happens.

Download a URL using Chrome as a fallback when fetch with binary:true and savePath cannot handle it. Try via:"page" for sites requiring page credentials or Referer. Chrome settings or policy may show a save dialog; explain that possibility and obtain permission before using this fallback during background work.

Attach a local file to the page (no extension can touch the OS file picker). Open the upload UI first. Auto-picks a matching file input; for editors that only accept drag-and-drop (X Article, Notion) pass dropSelector instead. Any size via a file input (browser reads the path); drag-drop caps at 48MB.

Extract structured data by CSS selector (lists, tables), or FIND BY TEXT with contains: every visible element whose text includes it, with a selector path and the ref it sits in — use it instead of eval to check "is X on the page", read a status label, or find a dropdown option. For lists pass html:true first to inspect markup, then extract: field → sub-selector, "@attr" for attributes, e.g. {title:".name", link:"a@href"}.

Your DEFAULT way to act — batch every step you can predict (on form wizards, nearly all). Each step is effect-checked, ONE snapshot returns at the end; every call you merge saves a full model turn. Stops early on no-effect / failure / submit-pay-delete controls. read {ref|find|selector, attr?} or {contains} brings an observation back mid-batch. Blocks (one level): repeat {steps,until,max} for pagination/load-more; if {cond,then,else} for optional banners; assert {cond} stops unless the page matches. cond = {urlContains|selectorExists|textContains, not} (OR-ed), or {ref|selector, checked|value|text} for one element. Inside repeat use find/selector, never ref. Elsewhere: ref until the page re-renders, find after.

Hand control back to the user for one step, then continue. For captcha, QR login, SMS/OTP, or any confirmation that should be a human decision. Brings the tab forward, shows a panel, highlights your targets, sends a desktop notification, and blocks until the user acts. Use it instead of retrying a step that needs a human. A "cancelled" result means the user said no: stop that task, do not look for another way to do the same thing.

One short sentence telling the user what you are about to do, shown on an on-page panel. Call before a multi-step task and whenever the plan changes. Instant, never blocks. Use the user's language.

Run a JS expression in the page. NOTE: fails on any site with a strict CSP (no unsafe-eval) — most large sites. Prefer query for data extraction. Isolated world: sees the DOM, not the page's own JS variables.

Site notes: APIs, walls, pitfalls from past sessions. Call {domain} before first acting on a site (no args = list sites). Notes may embed runnable ```act scripts — fill {{placeholders}} and run them instead of rediscovering. Learned something non-obvious or mapped a flow? Save the full note back via {domain, save}. Hints, never rules.

Reload the beat-browser extension (chrome.runtime.reload()) so an updated build takes effect — no manual chrome://extensions visit needed. ~2s; bridge auto-reconnects, tabs keep state. DISRUPTIVE: the extension is shared by every tab and every other agent session on this machine — anything mid-flight elsewhere gets cut. Only call right after installing/updating, or when doctor/a version-mismatch warning says to — never speculatively.