ipads-sai/mobiagent is an agent project that directly operates phone screens. A GUI agent is software that looks at a screen and presses buttons to get work done, and this project picked the phone as its target. The latest version on the GitHub releases tab is v1.0.1, and the changes included in that version are not given.
The structure splits into three axes. The MobiMind model family, the AgentRR acceleration framework, and the MobiFlow benchmark all sit in one repository. The README describes MobiAgent as a customizable mobile agent system. Actual execution is handled by a Python-based runner. That runner controls the phone with an Android app and ADB, the tool for driving an Android device from a computer by command, and MobiMind works on the smartphone task together with the runner. It supports user profile, experience, and behavior memory, and it also supports on-device inference, that is, running the model inside the device. The license is Apache-2.0 and the main language is Python. The repository was created on 2025-08-29, and the GitHub star count is 1921.
Still, it should be clear how far the evidence reaches. MobiFlow is a benchmark the repository ships itself, so no comparison against an external baseline is given, and no figures of its own are presented either. The conditions under which on-device inference holds, on which devices and with which models, are not given either. On a 10-point scale the only recorded score is 0.0 for momentum, and the novelty, technical, and quality scores are not given. So for now it is better to record this as a candidate phone GUI agent and not to judge its real performance until an external evaluation appears.