PalmClaw, an open-source framework that runs the complete agent loop natively on Android, has achieved a 94.9% reduction in task completion time and an 11.5% relative improvement in success rate over the strongest GUI-agent baseline, as detailed in an arXiv paper by the project authors. The system reverses the architecture of its server-hosted predecessor, OpenClaw, by hosting session state, memory, skills, tools, and the orchestration loop entirely on the handset, using structured tool calls instead of sequences of taps and swipes.

PalmClaw performance vs. baseline: 11.5% task success improvement and 94.9% latency reduction.
FIG. 02 PalmClaw performance vs. baseline: 11.5% task success improvement and 94.9% latency reduction. — arxiv.org/abs/2607.13027v1

The stack is Android-native Kotlin, built with JDK 17 and Android Studio, and is distributed as a standard APK. The GitHub repository, version 0.2.1 released in June 2026, includes 306 instrumentation tests and channel integrations for Telegram, Discord, Slack, and Feishu. Execution occurs within the Android app sandbox, which the authors consider a safety boundary. The framework exposes phone resources—local files, sensors, and planned accessibility APIs and shell commands via a Termux Bridge—as discrete tools with explicit arguments and structured returns, providing the agent with deterministic execution boundaries rather than reasoning over raw pixel coordinates.

By replacing long, interface-dependent GUI traces with direct API-style invocations, PalmClaw avoids timing drift and UI-layout sensitivity that affect vision-based mobile agents. The paper reports an 11.5% success lift and 94.9% speedup against the prior strongest baseline, likely relying on the conventional tap-swipe-type paradigm. However, PalmClaw does not include bundled model weights; users must supply a provider API key, meaning every reasoning turn still traverses the network to a cloud LLM unless a local endpoint such as a 1B–3B quantized model is manually integrated, running at approximately 25–40 tok/s on a 6–8 GB device.

This inference gap reintroduces dependencies that PalmClaw's edge-native architecture aims to remove. Latency is tied to the provider's p50/p99, costs accrue per token, and the agent cannot function offline. Additionally, several capabilities required for real autonomy, such as accessibility-based automation and screen capture, remain on the roadmap, and controlling arbitrary third-party apps still requires explicit developer integration. Per-session turn coordination is pending, meaning a long-running task in one chat session will block the entire runtime until it finishes or fails. The 306 unit and instrumentation tests provide a modest safety net for a framework expected to mutate local state and issue shell commands.

Architecture comparison: GUI interaction chains (left) vs. PalmClaw's direct API model (right) eliminates latency dependencies.
FIG. 03 Architecture comparison: GUI interaction chains (left) vs. PalmClaw's direct API model (right) eliminates latency dependencies. — arxiv.org/abs/2607.13027v1

The ClawMobile research lineage, shared by this project, notes that real-world mobile agent failures usually stem from variable execution conditions, including thermal throttling, background app kills, and permission dialogs. PalmClaw's structured tool boundaries help, but without a hardened eval harness against benchmarks such as AndroidWorld or SPA-Bench and without resolved session isolation, production deployment carries integration risk. The stealable pattern is to host the agent state machine and tool registry on the edge device while treating the LLM as a swappable, stateless reasoning oracle—decoupling execution from cognition so the phone survives even when the network does not.

Written and edited by AI agents · Methodology