PokeClaw is a proof-of-concept Android app that uses Gemma 4 to autonomously control a phone - tapping buttons, opening apps, navigating interfaces - entirely on the device, with no internet connection required. The developer posted a working demo on April 6, 2026, calling it the first app of its kind.
The demo works by having the model take screenshots of the phone screen, analyze what's visible, decide what action to take next, and execute touch interactions. All of this runs using Gemma 4 (Google's open-weight model, meaning the model weights are publicly downloadable and can run on your own hardware), directly on the phone's processor - not on a remote server.
Running on the Phone, Not the Cloud
Current AI phone assistants route complex requests through remote servers. That means your screen content, your queries, and your actions are sent to and processed by third-party infrastructure. An on-device model that completes tasks locally never sends that data anywhere - which matters for anyone using their phone for sensitive work or communications.
For the on-device approach to work, the model needs to be small enough to fit in phone memory while still being capable enough to look at a screenshot and make a reasonable decision about what to tap next. Gemma 4 hits that threshold. A year ago, fitting a model with that level of visual reasoning capability onto a phone chip would have required significant hardware compromises.
What It Doesn't Do Yet
PokeClaw is a demo, not a finished app. Autonomous phone control is notoriously brittle - UI layouts change, apps update, and the same task looks different across devices and Android versions. The current implementation handles a limited set of tasks and almost certainly fails in ways a production app can't afford.
The technical signal here is that this architecture works at all. Gemma 4 is Google's April 2026 model release, with multimodal capability (it processes both screenshots and text, not text alone). Running it on-device for real-time screen control establishes that the building blocks are in place. Whether PokeClaw matures into a practical tool or stays a community demo, the underlying pattern is now proven.