TencentQQGYLab/AppAgent

Operates Android apps through vision-based perception and simplified action primitives (tap, swipe) without requiring backend access, making it applicable across diverse third-party applications. The framework employs a two-phase approach: an exploration phase where the agent builds a knowledge base either through autonomous navigation or human demonstrations, followed by a deployment phase that leverages this documentation to execute complex tasks. Supports multiple multimodal LLM backends including GPT-4V and Qwen-VL, with integration via Android Debug Bridge (ADB) for real devices or Android Studio emulators.

6,582 stars. No commits in the last 6 months.