Start a browser session (cold-launch real Chrome).

Attach to an existing Chrome via CDP (use after manual login on a Cloudflare-walled site).

Stop the browser session.

Daemon + session status.

Fast pre-check (DNS, optional HTTP HEAD) for a URL before go commits to a 30s navigation timeout. Catches typos and dead domains in ~50ms instead of 30s.

Toggle the daemon's per-verb DEBUG audit log at runtime. Useful for non-trivial runs where you want a full call trail; the toggle's response returns the daemon-log path to tail.

USE THIS WHEN WebFetch/WebSearch/curl GOT BLOCKED — a 403, a Cloudflare/DataDome "Just a moment..." challenge, a login/paywall, or an empty "JavaScript is required" shell: don't report failure, use this — patchright clears most such walls cold. ONE-CALL 'look at this URL', TEXT-FIRST. Does a DNS pre-check → auto-start a session if needed (headless) → navigate → extract the text. It does NOT screenshot by default: read and navigate with the returned text/selectors, and a screenshot is captured only as a FALLBACK when the page yields no usable text (canvas/image/blank SPA). This is the 80% case of 'just show me what's on this page' — use it instead of separate start / navigate / read calls unless you need multi-step interaction. Pass screenshot='always' to force a screenshot, 'never' to suppress even the fallback; when one is captured it comes back as a viewable image, not base64 text. Without an explicit session it runs on an OFF-BUDGET transient ephemeral session (0.7.0) — it never competes with persistent sessions for a slot and is auto-deleted afterward, so it won't touch your 'default' session.

ONE-CALL verification gate. Assert the page reached an expected state — composes element-state / page-text / URL checks plus a native challenge-wall check into a single {passed, failures[]} verdict. Use after an action to confirm it landed (or that you got soft-blocked) instead of stitching several wait and read calls together. Every check is optional.

Navigate to a URL in the stealth browser, then read the page with a content tool. Reach for this when WebFetch/curl was blocked — a 403, a Cloudflare/DataDome "Just a moment..." challenge, a login/paywall, or a JavaScript-only page that returns no usable HTML — or when the page needs a real browser.

Browser back.

Browser forward.

Reload current page.

Get current URL.

Get current page title.

Get inner text (whole page or a target — @eN, @text:Foo, @label:Email, CSS, etc.). After a click or form step, pass settle=true: it waits for the page to stop changing, then reads, and returns the url too. Use that instead of a timed wait in page script.

Get HTML (whole page or a target — @eN, @text:Foo, @label:Email, CSS, etc.).

Clean, LLM-ready Markdown of the page (or a target subtree) — boilerplate (nav/footer/aside/scripts) stripped, headings/links/lists/code preserved. A drop-in for Crawl4AI/Firecrawl-style scraping on the AUTHENTICATED pages those stateless tools can't reach. Returns markdown TEXT (never a base64 screenshot), capped by max_chars. mode: markdown (default) | links (deduped {url,text}, absolute post-hydration URLs) | assets ({url,type,rel?}, data: dropped) | main (main-content only via density scorer).

Declarative STRUCTURED extract: pass a {name: selector} map and get back ONE JSON object of values in a single call — against our real authenticated Chrome DOM (a login-walled / SPA page stateless scrapers can't reach). Grammar: name[]=array, sel@attr=attribute, sel@html=innerHTML, bare=text. Optional target scopes selectors to a subtree. Returns {fields, matched, misses, errors} — misses/errors let you fix a selector without re-reading the page. Reads text/attr/innerHTML only, never input values.

Structured map of every on the page (plus a formless-controls group for SPAs) — against our real authenticated Chrome DOM. Each field carries a ready-to-use locator string (pipe straight into fill/click), plus type, name, label, required, options and checked/filled state. A free-text field's typed value is withheld unless values=true, and even then it's redacted when a type/name/autocomplete heuristic flags it as a password/credential/payment secret (best-effort — don't pass values=true on untrusted pages). Returns {forms, count}.

List EVERY element a target resolves to, so an ambiguous locator can be disambiguated instead of failing strict mode. Returns {target, count, candidates:[{index, tag, role, name, text, bbox}], truncated}. Act on one with click/fill/type/hover passing index=N.

Get an HTML attribute value from an element.

Get the current value of an input/textarea/select.

Evaluate a JS expression in the page (isolated context per Patchright default — window.X = ... mutations are NOT visible to page JS; use eval_handle to retain references).

Snapshot actionable elements and assign @eN refs (uses Playwright aria_snapshot mode='ai').

Diff current snapshot vs the previous one.

Click an element. Target: an @eN ref from map_compact, visible text (Add to cart), @text: / @label: / @role:button[name=Pay], or a CSS selector. If the page covers or disables the target the error says what is in the way.

Double-click an @eN ref or CSS selector.

Focus an element (without clicking).

Pick an option on a by value/label/index. Also works on a custom dropdown: pass the button that opens it and the option's label.

Clear an input and fill it with text. With use_secret, value comes from the encrypted vault — and is only written into a page on the secret's own site (https, same host or a subdomain, or the entry's origins); otherwise refused.

Type text key by key into an element. Use this rather than fill when the field offers suggestions as you type, such as an address box.

Hover over an element.

Press a key on a specific element (e.g. Enter on @e3).

Press a key combination (e.g. 'Control+a', 'Enter').

Check a checkbox / radio.

Uncheck a checkbox.

Scroll the page or a target element into view.

Check element state (visible/enabled/checked/...).

Capture a screenshot. With no path it returns a viewable image (not base64 text); with a path it saves the PNG and returns the path. With tiles=true it slices a full-page capture into fixed-height tiles written to DISK (returns their paths + count, never base64) — for layout-heavy pages a vision-capable caller then reads tile-by-tile. Prefer reading the page text/selectors first — only screenshot when you actually need the pixels.

Wait until an element reaches a state. Takes the same targets as click.

Wait until an @eN ref reaches a state.

Wait until the URL matches a glob/regex.

Wait for a page load state (load/domcontentloaded/networkidle).

Wait until a JS expression returns truthy.

List all open browser tabs.

Open a new tab and switch to it.

Switch to tab by index.

Close the current tab; the next tab becomes active.

Locate elements by semantic strategy (text/label/placeholder/role/testid/xpath/alt/title/css).

Count matching elements for a selector or @eN ref.

Replace the page HTML wholesale via page.set_content.

List all live frames with name + url + depth + active flag.

Switch active frame by name or URL substring (omit both to clear).

Mouse control at pixel coordinates.

Set files on an input[type=file] target.

Arm the NEXT native file chooser of the active tab with files, for upload buttons that have no input[type=file] in the DOM. Then click the button (any verb). One shot; expires after timeout_s. Choosers are intercepted only while armed. Check with upload_status.

Whether the armed file chooser was filled (consumed), plus how many choosers were swallowed (no arm left) or foreign (from another tab while armed).

Save the current page as PDF.

Briefly outline an @eN ref or selector for visual debugging.

Screenshot with @eN bounding-box overlays (vision-LLM friendly).

One-line-per-element rendering of the snapshot (token-efficient): @eN role "name" [state…], preserving state ([checked]/[disabled]/[expanded]/[level=N]). interactive=only actionable roles; bbox=append real bounding_box() coords (bbox=x,y,w,h) — genuine layout geometry.

Plan a verb + target for an intent without executing.

Do one action from a short instruction, without naming a selector. It matches the words against element names, with no reasoning, so use the page's own wording: click Add to cart, fill Email with a@b.co, select Victoria in State, tick Accept terms. One action per call. Check executed and error.

List all profiles and the active one (alias of session_list).

Create a new named profile (alias of session_new).

Set the active profile (alias of session_use).

Delete a profile directory (alias of session_delete).

Create a new on-disk session/profile dir (does NOT launch Chrome).

List every on-disk session + which are currently running.

Set the active session (persisted to active-session file).

Alias for session_use.

Stop Chrome for one session (profile dir preserved).

Stop Chrome for every running session.

Delete a session's profile dir on disk (not active/default).

Acquire/renew an exclusive advisory lease on a running session (0.7.0). Returns a token; present it as the lease arg on subsequent calls so other clients get a 'busy' error instead of clobbering the shared page.

Release a session lease — present the token via the lease arg, or force=true to break it (operator override).

Report the lease state for a session (never returns the token).

Housekeeping — reclaim disk from stale profile dirs, superseded Chrome cache dirs, leftover Chrome lock files, regenerable caches, and the daemon log. DRY-RUN by default; pass apply=true to actually delete. Never touches the default/active/running sessions. Returns a per-category {count, bytes} report.

Enable humanlike mouse (Bezier paths, gaussian dwell, sin scroll) for current session. With a recorded swipe library installed, the wheel replays the operator's swipes on one device per session.

Disable humanize for current session.

Report whether humanize is on for current session.

Opt-in AMBIENT activity between verbs (default off): while the session idles, emit low-rate human-plausible pointer drifts + rare reading scrolls (never clicks/types/selects/navigates, yields to every verb, stops horizon_s after the last verb or at idle-freeze). Targets are best-effort: it avoids elements that look interactive but can't see script-attached hover handlers, so a page may still observe a mouseover. Fills the zero-mouse-events silence behavioural scorers flag. mode=on|off|status.

Find a UI element by verbal description (via Claude vision) and click it. Fallback for canvas/Flutter/Unity pages where AX-tree fails. Requires [llm] extra.

Return coords + confidence for a described element (no click).

vision_click the described field, then type the given text.

Return cumulative vision API usage (calls, cache_hits, tokens, cost_usd) for session.

Drop the on-disk vision coords cache.

Report today + lifetime vision spend vs VIBATCHIUM_VISION_MAX_*_USD caps. reset='today'|'lifetime'|'all' clears the spend log.

Heuristically dismiss cookie/consent/newsletter banners.