36 typed browser tools.

Navigate the active context to a URL. Optionally waits for load + page-settled before returning.

Return the optimised DOM snapshot for the current page (semantic JSON, ~93.1% smaller than Chrome AX).

Click by CSS selector. Auto-scrolls into view, action-locks during the press, verifies the click landed.

Focus a selector and type into it. Triggers input + change events. Falls back to keyboard input where IME interaction is needed.

Capture a PNG screenshot of the current viewport or a clipped element. Returns base64 + dimensions.

Scroll the page by a delta or to a target selector. Honours overflow containers, not just window scroll.

Allocate a fresh browser context from the pool. Optional citizen id pins identity; otherwise a Nomad session is used.

Release the active context back to the pool. Recorded in the lifecycle audit log.

Return the full accessibility tree for screen-reader-style reasoning. Filters injected DOM via the Phase 1 filter.

Click by durable reference id from a prior snapshot — survives minor DOM churn between turns.

Type into a referenced element. Same durability story as click_ref.

Hover over a referenced element. Useful for menus, tooltips, and lazy-rendered surfaces.

Wait for one of: element visible, text appears, network idle, DOM stable, URL contains. Configurable timeout and poll interval.

Search interactive elements by visible text / aria-label / placeholder. Returns positions, durable refs, and disambiguation hints.

Assert state after an action: exists, visible, checked, value, text, url, title. Emits a typed pass/fail trace row.

Take before / after screenshots and return a structural diff. Confirms an action actually changed the page.

Block until readyState=complete, no pending images, no in-flight network for N ms, and DOM mutation count below threshold.

Choose from a <select> by value or visible text. Fires change so frameworks update model state.

Fill multiple fields at once. Triggers correct input/change events per field type. Reports per-field success.

Return URL, scroll position, form / button / link counts, modal visibility, and current focus element.

Press a key or chord (Ctrl, Shift, Alt, Meta + key). Honours focus context and IME state.

Select-all + delete on a focused input. Works on contenteditable surfaces too.

Collect HTML5 validation errors, CSS error states, and ARIA error messages from the active form.

Click with cursor trajectory (Bezier path), arrival overshoot, and timing jitter calibrated to real-user telemetry.

Smooth scroll with easing curves, mid-scroll pauses, and direction reversals. Avoids the linear-scroll fingerprint.

Per-character delay variance, occasional corrections, and shift-modifier timing that matches human typing patterns.

Screenshot plus labelled interactive elements with durable object ids. Pairs with click_label for vision-driven agents.

Click by @N label returned from annotated_screenshot. Lower hallucination than CSS selectors for vision models.

Look up a credential for the active citizen via the credential provider. Returns metadata only, never plaintext.

Inject a stored credential into a targeted form field via Page.secureSetInputValue. Plaintext never crosses the JS boundary.

Start an audio capture session on the active page. Requires the AudioCapturer extension.

Stop an active audio capture session and finalise the segment.

Drain captured audio bytes from an active session for downstream transcription.

Enumerate Android devices visible to the mobile bridge. Returns device id, model, and Android version.

Start a mobile bridge session against a device id. Returns a CDP endpoint usable by the same toolbelt.

Stop a mobile bridge session and release the device.