Overview

PokeClaw, also called PocketClaw, is an open-source Android application that enables fully local AI-powered phone control. Built rapidly by a solo developer shortly after the release of Google's Gemma 4 model, it allows users to give natural language instructions to an on-device AI agent that can navigate the phone, interact with apps, and execute complex tasks — all without sending any data to the cloud.

"Poke" is short for "Pocket," emphasizing its portable, private, on-device nature.

Key Features

Fully On-Device AI: Runs Google's Gemma 4 model locally using LiteRT-LM runtime. No internet connection, no API keys, and no recurring costs after initial model download.
Phone Automation Agent: The AI can understand instructions and perform real actions such as opening apps, tapping elements, typing, scrolling, and handling multi-step workflows.
Privacy-First Design: All processing happens on your Android device. Ideal for users concerned about data privacy and security.
Open Source: Fully transparent codebase available on GitHub. Community contributions and customizations are welcome.
Task Planning & Execution: Supports complex, multi-step tasks with planning capabilities powered by the local LLM.
Easy APK Installation: Download the latest signed APK directly from the GitHub releases.

Technical Details

Model: Google's Gemma 4 (optimized for on-device inference)
Runtime: LiteRT-LM (formerly known as TensorFlow Lite or similar on-device LLM runtime)
Platform: Android (native app)
Architecture: Combines screen understanding, action prediction, and execution in a closed-loop agent system
Development: Built from scratch in roughly two days as a prototype, now evolving with community input

Installation

Visit the GitHub repository
Go to the Releases section
Download the latest APK
Install on your Android device (enable "Install from unknown sources" if needed)
Download the Gemma 4 model (handled within the app or via provided instructions)

Note: If updating from an older debug build, uninstall the previous version first before installing the new signed APK.

Use Cases

Daily personal automation (e.g., "check my messages and reply to urgent ones")
Accessibility assistance for users with motor or visual challenges
Privacy-sensitive workflows where cloud-based agents are unacceptable
Experimenting with local LLM agents on mobile hardware
Automating repetitive tasks like form filling, navigation, or content curation directly on the device

Comparison & Context

PokeClaw is frequently compared to cloud-based or hybrid phone agents (including projects like OpenClaw). Its main advantage is true on-device operation, offering lower latency for supported tasks, complete privacy, and zero ongoing costs.

PokeClaw