ClawGUI: A Unified Open-Source Framework for Training, Evaluating, and Deploying GUI Agents

Summary

ClawGUI is an open-source framework that addresses the full-stack infrastructure gap limiting progress in GUI agents—AI systems that drive apps through visual interfaces via taps, swipes, and keystrokes rather than programmatic APIs, reaching long-tail applications that CLI-based agents cannot. The framework consists of three integrated components: ClawGUI-RL, the first open-source reinforcement learning infrastructure for GUI agents; ClawGUI-Eval, which enforces standardized evaluation pipelines across 6 benchmarks and 11+ models; and ClawGUI-Agent, which deploys trained agents to Android, HarmonyOS, and iOS through 12+ chat platforms. As a demonstration, the end-to-end trained ClawGUI-2B model achieves a 17.1% success rate on MobileWorld GUI-Only. The paper is authored by Fei Tang, Zhiqiong Lu, Boxuan Zhang, Weiming Lu, Jun Xiao, Yueting Zhuang, and Yongliang Shen, and is available on arXiv (2604.11784).