Move the mouse cursor. Coordinates are logical points, origin top-left of the main display. Compute them from the screenshot tool's bounds_pt and px_per_pt metadata.

Click at (x, y), or at the current cursor if omitted. Coordinates are logical points, origin top-left of the main display. Compute them from the screenshot tool's bounds_pt and px_per_pt metadata. After clicking, take a screenshot to verify the result.

Press, drag along path, release. Coordinates are logical points, origin top-left of the main display. Compute them from the screenshot tool's bounds_pt and px_per_pt metadata.

Scroll at (x, y) (moves there first) or at the current cursor. Positive dy scrolls up, negative down, in wheel ticks unless pixels=true.

Type text into the focused element using layout-independent unicode injection. Click the target field first.

Press a key or shortcut chord, e.g. 'Return', 'cmd+s', 'cmd+shift+p', 'F5'. Use keyboard_type for regular text.

Capture the screen (or a region, in points). The response includes bounds_pt and px_per_pt: point = bounds_pt.origin + pixel / px_per_pt. Take a screenshot before any coordinate action.

Re-orientation: cursor position, displays, last screenshot metadata, kill-switch status, and recent action history.

Wait for a condition: {type: 'duration', ms} | {type: 'text_present', text, region?} | {type: 'screen_stable', quiet_ms}. Timeout returns met=false (an answer, not an error). text_present requires an exact (case-insensitive) substring match, not fuzzy matching.

OCR the screen (or a region, in points) and return boxes matching text. Each match has a center you can pass directly to mouse_click. Re-observe rather than act when confidence < 0.5. By default this matches fuzzily (case-insensitive substring either direction, or a similarity ratio >= 0.8); pass fuzzy=false for exact substring matching. This differs from wait's text_present condition and verify's text_present expectation, which always require an exact substring match.

Check an expected outcome after acting. expect = {type: 'text_present'|'text_absent'|'region_changed'|'region_unchanged'|'cursor_at'|'window_present'|'window_gone'|'clipboard_contains'|'all_of'|'any_of', ...params, children?}. region_changed/unchanged need baseline_screenshot_id from an earlier screenshot response. text_present/text_absent require an exact (case-insensitive) substring match, not fuzzy. clipboard_contains never returns clipboard content, only matched/clipboard_len.

Read the clipboard (sensitive: may require user confirmation). Refused while a password field has focus.

Set the clipboard to text or a base64 PNG.

Paste text into the focused app via clipboard + Cmd+V, then restore the previous clipboard. Preferred over keyboard_type for long text.

List windows (optionally filtered by app bundle id or name). Returns window_ref handles for window_focus/window_manage.

Focus (raise) a window by window_ref, or by app and/or title_match. Stale refs are re-resolved by pid + fuzzy title.

move/resize/minimize/unminimize/maximize/close a window. close may trigger 'Don't Save' dialogs and needs confirmation under the default policy.

Launch an app by bundle id (preferred) or name; activates it if already running. wait_for_window waits for its first window.

Quit an app gracefully; force=true force-terminates (sensitive, needs confirmation under the default policy).

List running apps and the frontmost one.

Accessibility tree for an app (frontmost if omitted): roles, titles, values, clickable regions in points. Ground truth where apps expose it; use find_text where they don't.

Run up to 20 PRE-DECIDED acting steps in one call (click, type, press...), each optionally gated by a guard condition (same schema as wait). A failed guard halts the sequence and returns evidence. Use only when you already know every step; this is a macro, not a planner.