Move mouse to specified screen coordinates. Args: x: Horizontal position (0 = left) y: Vertical position (0 = top) relative: If True, moves relative to current position (default: False) Returns: dict: { 'success': bool, 'error': str (if failed) } Examples: move_mouse(100, 200) # Moves to absolute x=100, y=200 move_mouse(10, 10, relative=True) # Moves 10px right and down

Simulate left mouse click at current position. Returns: dict: { 'success': bool, 'error': str (if failed) } Example: click_mouse() # Clicks at current cursor position

Perform drag operation between coordinates. Args: x1, y1: Start position x2, y2: End position Returns: dict: { 'success': bool, 'error': str (if failed) } Example: drag_mouse(100, 100, 200, 200) # Drags from (100,100) to (200,200)

Scroll vertically (positive=up, negative=down). Note: Each unit represents one notch on the scroll wheel (120 units = high-definition scroll). Typical values range from 2-3 to 5-10 for normal scrolling.

Capture the screen, with measurement rulers drawn on the result.

Compare two images using a VLM. Requires a configured VLM provider.

Analyze a screenshot with a VLM. Requires a configured VLM provider.

Execute system actions with chaining support. Handles both single actions and chained sequences. Args: action: Action string in format: Single: "prefix:params" (e.g. "click:100,200") Chain: "chain:action1;action2" (e.g. "chain:click:100,200;type:hello") Supported Actions: type:text - Type text press:key - Press key click/click: - Click at current position (both formats supported) move_to:x,y - Move to absolute coordinates (default) move_to:rel:x,y - Move relative to current position drag:x1,y1:x2,y2 - Drag between points scroll:amount - Vertical scroll (positive=up, negative=down) Note: Each unit = 1 scroll notch (120 = high-def scroll). Typical: 15-120. scroll:horizontal:amount - Horizontal scroll Note: Each unit = 1 scroll notch (120 = high-def scroll). Typical: 15-120. Returns: dict: {'success': bool, 'error': str}

Declared as a dict because that is what every code path returns. It was
annotated -> bool while returning dicts, and a strict MCP client rejects the
structured content against the declared schema, so the tool was unusable
from a conforming client.

Example: execute_action("click:100,200") execute_action("chain:click:100,200;type:hello;press:Enter")

Capture then analyze the screen. Requires a configured VLM provider.

Report which backends this machine offers, and which were selected.

Call this first when a capture or an input action fails: it says what is installed, what the compositor advertises, and what that adds up to.