MobileAgent Android
An Android app implementing the Mobile-Agent autonomous agent framework, letting a vision-language model such as GPT-4o or Claude see a phone's screen and operate it by tapping, swiping and typing to complete a task described in natural language.
MobileAgent Android is an Android-native implementation of the Mobile-Agent framework. It captures screenshots, detects interactive UI elements through the Android Accessibility API, and feeds annotated images to a vision-language model, which plans and issues actions such as tap, long press, swipe, text input, system button presses and a final answer or completion signal.
Its agent loop follows a four-phase architecture of a manager for planning, an executor for acting, a reflector for evaluating results and a notetaker for memory, running for up to a configurable number of steps with automatic error recovery and replanning. A floating status window shows live progress, and a detailed execution log records each step's prompts and responses; the interface is available in English and Simplified Chinese.
It supports OpenAI-compatible APIs and Anthropic's Claude for the underlying model, and setup requires granting Accessibility Service, overlay, notification and battery-optimization-exemption permissions. It is built with Kotlin and Android Studio, and is MIT licensed.