Mobile Device MCP Server

Agents can discover available devices, capture JPEG screenshots, and inspect a normalized, flat UI element list before interacting with a screen. Tree queries support search and result limits; off-screen elements are omitted unless requested. Interaction tools cover coordinate-based taps, double taps, long presses, swipes, text entry into the focused element, and hardware or navigation buttons. App management includes listing installed apps, launching or stopping them, and installing host-side files with replacement of existing installs.

For scripted interactions, run_code executes JavaScript through Android UIAutomator or iOS XCUITest bindings. Android uses Rhino, while iOS uses JavaScriptCore; each call receives a fresh sandbox, and runaway scripts are stopped. On first use, the server installs its device driver, assigns a port, and starts an on-device HTTP server. Subsequent tool requests use localhost HTTP and a bearer token specific to that device.