Researchers present PalmClaw, an open-source agent framework that runs natively on mobile devices to manage sessions, memory, skills, tools, and the agent loop directly on the phone. Unlike existing agents that rely on GUI actions like tapping and swiping, PalmClaw exposes device capabilities as explicit tools with structured results and defined execution boundaries.
- The framework manages agent components locally on the device rather than relying on external servers.
- It allows agents to access mobile hardware and applications directly through explicit tool calls.
- Experiments show an 11.5% relative improvement in task success rates compared to the strongest baseline.
- Completion time is reduced by 94.9% over the baseline, with lower setup requirements.
This design enables more controlled and efficient agent execution on mobile hardware by eliminating the need for interface-dependent sequences.