Scout

Get a compact accessibility tree showing all interactive elements.

Label all interactive elements with numbers and return their selectors/info. By default returns only the element list (compact). Set include_image=true to also get the screenshot with labels drawn on it. NOTE: label numbers are per-call only — after any DOM mutation (re-render, modal open, route change), the same label may point to a different element. Pair click_label with the most recent annotated_screenshot in the same turn; for stable identity across mutations use the returned selector.

Extract global app state (Redux, Next.js, Nuxt, Remix, SvelteKit, Gatsby, Astro, Alpine, HTMX).

Minimal accessibility scan grouped by impact (critical/serious/moderate/minor). Covers image-alt, button-name, link-name, label, duplicate-id, html-has-lang, landmark-one-main. Zero dependency — for full WCAG coverage run axe-core in your test pipeline.

Check that the given text appears on the page in one round-trip. If selector is set, searches only that element's text; otherwise searches body innerText. Returns ok=true with a surrounding snippet on hit, ok=false with a short page-tail snippet on miss. Replaces extract+string-match for post-action verification.

Auto-detect repeating patterns (product cards, search results, list items) and extract structured data. No selectors needed.

Execute multiple actions in a single call. Avoids repeated round-trips. Actions: click, type, fill_form_semantic, wait, scroll_to, click_label. Continues on error.

Check how ready the page is for interaction. Returns a 0-100 score, pending XHR count, skeleton/spinner presence, and suggestions for what to wait for.

Click an element by CSS selector. Set wait=true for navigation clicks.

Click the element identified by a stable node_handle from a previous annotated_screenshot. Unlike click_label, the handle survives DOM mutations that would shuffle label numbers. Returns a stale_handle error when the element is no longer in the DOM — re-annotate and retry.

Click an element by its label number from annotated_screenshot.

Click an element by its visible text. Resolution: aria-label exact match, then button/link text match, then closest interactive ancestor of a matching text node. Pass role="button"|"link" to disambiguate when text appears in both. Cheaper than annotated_screenshot for the common 'click that thing I can see' case.

Close a named tab. Cannot close the currently active tab.

Compare content between two named tabs. Returns what's different, what's only in one tab.

Extract component state/props from any framework (React, Vue, Svelte, Angular, Alpine, Lit).

Change browser settings without restarting. Use headless=false to see the browser window. Set allow_private_ips=true for local-dev workflows (localhost, 127.0.0.1, private IPs). Pass fresh=true to force-kill the current page and start clean (use after observing session_dead in status).

Get captured console.error / console.warn messages plus recent network 4xx/5xx failures. Auto-installs lightweight network observers, so failures recorded after the first call surface here without an explicit enable_network_capture.

Clear cookies. With no name, drops all cookies via Network.clearBrowserCookies. With a name, deletes only that cookie (optionally scoped by domain/path). Use this when a backend has been restarted and the previous session token is invalid.

List all cookies for the active page (names + flags only — values are redacted). Useful for diagnosing stale-session issues after a backend restart.

Set a cookie on the active page. Useful for restoring sessions or seeding state.

Check if the page is a login wall, paywall, or CAPTCHA. Returns type, confidence, and reason.

Check if a modal, dialog, popup, or overlay is currently visible. Returns its title, text, buttons, and inputs.

Detect which frontend frameworks are active on the page.

Discover form fields with their labels, types, and CSS selectors.

Auto-dismiss cookie consent banners. Tries common selectors and text patterns (Accept, Agree, Got it, OK). Returns whether a banner was found and dismissed.

Dispatch a DOM event on an element. Useful for triggering SPA event handlers. Special cases: event_type='submit' on a <form> (or any element inside one) calls form.requestSubmit() — the only reliable way to fire Vue @submit.prevent / React onSubmit and run HTML5 validation. event_type='click' invokes el.click() so the default action runs (e.g. submitting the owning form when target is type=submit).

Double-click an element.

Drag an element and drop it on another element.

Start capturing network XHR/fetch responses. Patterns filter by URL substring.

Extract text content from a single element.

Extract text from all elements matching a selector.

Extract structured data from an HTML table (headers + rows).

Recent network requests with status >= 400 (4xx/5xx). URL + method + status + response body snippet. Use after a form submission that silently failed.

Fill multiple form fields at once.

Fill form fields by label or name. Strings fill text inputs; booleans toggle checkboxes; for radios pass the label/value as a string. Each result includes the value re-read after dispatching input/change events plus a warning when framework binding (Vue v-model / React onChange) didn't pick up the change.

Find the interactive element at given pixel coordinates. Returns the smallest element containing that point, with CSS selector and text. Use after hybrid_observe.

Set focus on an element, triggering :focus CSS state.

Check if an element exists on the page.

Hover over an element to trigger CSS :hover states, tooltips, and dropdown menus.

Vision+DOM hybrid mode: returns a clean screenshot (no labels) plus bounding boxes for all interactive elements. Use find_by_coordinates to select elements by pixel position. Set include_image=true to get the base64 screenshot.

Pre-install the submit tracker so the next submit event records defaultPrevented + XHR count. Call after navigate, before triggering a form submit, if you plan to read submit_outcome afterwards.

List all open tabs with their names, URLs, and titles.

Get a compact markdown representation of the page. Ideal for LLM processing.

Navigate to a URL. Returns page title and URL.

Get captured network requests/responses including request bodies (POST/PUT/PATCH) and response bodies (max 32KB each, truncated if larger). Includes recent buffered requests so you can inspect traffic even if capture was enabled late. If the list is empty and capture was never enabled, the response includes a hint pointing at enable_network_capture.

Rolled-up view of captured network traffic. Returns total count, bucketed by status class (1xx/2xx/3xx/4xx/5xx/0), every request with status >= 400 inline, and the pending count. Single call replaces enable_network_capture → act → failed_requests for the common 'did this action succeed at the network layer' question.

Get a structured snapshot of the current page. Optionally pass url to navigate first.

Return only page changes since the last observation.

Observe a subset of the page. Limit by landmark roles (nav/main/footer/...) and/or cap text length and element counts. Use on listing pages where the unscoped observation eats more tokens than the task needs.

Observe the page within a token budget.