PalmClaw brings LLM agents onto mobile devices

LLM agents have expanded from response generation into task execution, but deployment has largely remained tied to desktop and server environments. Mobile devices offer ubiquitous, data-rich, sensor-laden computing environments, yet existing mobile agent approaches often rely on inefficient GUI actions and struggle to access device capabilities directly.

PalmClaw runs natively on mobile phones, managing agent loops, memory, skills, and tools on the device. It exposes device capabilities as tools with explicit arguments and structured results, allowing agents to bypass indirect GUI interaction and operate within clearer execution boundaries.

Experiments for the framework showed an 11.5% relative improvement in task success rates compared with the strongest baseline and a 94.9% reduction in task completion time. The lower setup burden and faster execution position PalmClaw as a path toward more responsive on-device AI applications.