PokeClaw offers the first on-device AI capable of autonomously controlling an Android phone, utilizing the Gemma 4 model without requiring cloud connectivity, API keys, or recurring costs. This project targets users and developers seeking private, offline mobile automation, providing a closed-loop system where the AI pipeline operates entirely on the user's device, enabling context-aware actions and task execution.

How It Works

PokeClaw leverages Gemma 4, a 2.3B parameter LLM, running locally on Android devices via the LiteRT-LM runtime. The system captures the phone's screen as text, which the LLM interprets to select and execute a suite of generic tools (e.g., tap, swipe, type, open app, send message). This process creates a closed loop: Phone -> LLM -> Phone. A novel "Skills" system, inspired by Claude Code, allows for reusable workflows built from these generic tools, enabling reliable automation for complex tasks that small local models might otherwise struggle with.

Quick Start & Requirements

• Primary install: Download and install the provided APK.
• Prerequisites: Android 9+ (12+ recommended), arm64 architecture, 8 GB RAM minimum (12 GB+ recommended), ~3 GB free storage for model download. No root required.
• Setup: Grant accessibility permission upon first launch; the model (~2.6 GB) downloads automatically.
• Links: Landing Page: https://agents-io.github.io/PokeClaw/

Highlighted Details

• 100% On-Device: Operates entirely locally, ensuring privacy and eliminating cloud dependency, API keys, and associated bills.
• Context-Aware Automation: Demonstrates advanced capabilities like reading full conversation history for contextually relevant auto-replies on messaging apps.
• "Skills" System: Predefined, extensible workflows composed of generic tools enable reliable task automation, acting as "training wheels" for the LLM.
• Hardware Adaptability: Functional on CPU-only inference (though slow, ~45s warmup on budget devices) and significantly faster on modern chipsets like Google Tensor G3/G4 or Snapdragon 8 Gen 2/3.

Maintenance & Community

This project is developed by a solo developer, with active updates indicated by recent changelogs (e.g., v0.3.2 on 2026-04-07). The developer encourages users to report bugs and propose fixes via GitHub Issues. No dedicated community channels (like Discord or Slack) are listed.

Licensing & Compatibility

• License: Apache 2.0.
• Compatibility: Compatible with commercial use under Apache 2.0 terms. Forks must be renamed before distribution, and the "PokeClaw" trademark may not be used to endorse derived products without written permission.

Limitations & Caveats

PokeClaw is explicitly described as a 2-day open-source prototype, not a polished consumer application, and contains numerous issues. Local LLMs are inherently less capable than their cloud counterparts, leading to rough edges. Performance is highly dependent on device hardware, with CPU-only inference being notably slow. The minimum 8 GB RAM requirement is tight and may lead to crashes on some devices.