ClawGUI
A research framework covering the full lifecycle of GUI agents: online reinforcement-learning training, standardized evaluation across benchmarks, and deployment to control real mobile devices by natural language.
ClawGUI is a research framework for GUI agents that covers online reinforcement-learning training, standardized evaluation, and real-device deployment as one pipeline, addressing the three problems together: an environment to train an agent online, benchmarks to measure what it learned, and a production system to deploy it.
Its modules are ClawGUI-RL, which trains agents online in parallel Docker environments or on real Android devices using GiGPO with process reward modeling for step-level rewards; ClawGUI-Eval, which measures models against six benchmarks with reproduction close to official published results; ClawGUI-Agent, which controls mobile devices from natural language through more than a dozen chat platforms; ClawGUI-Skills, which lets agents evolve reusable skill packages from experience without further training; and ClawGUI-APP, which runs the full stack directly on one Android phone.
A 2B-parameter model trained end-to-end with this pipeline is released alongside the framework's paper.