PokeClaw - An Open-Source Fully Local Offline Mobile AI Agent App

What Is PokeClaw

PokeClaw (full name PocketClaw) is an open-source Android AI automation app independently built by developer Nicole. It is the first mobile AI agent to achieve fully local offline operation, performing inference directly on-device via the LiteRT-LM runtime based on the Google Gemma 4 model, and can autonomously control a phone without internet access, an account, or an API key. By reading the screen UI tree through Android's accessibility service, the AI can perform actions such as tapping, swiping, entering text, and sending messages, achieving a true closed-loop agent. All data processing is done locally, fully protecting privacy; it can also optionally connect to cloud models such as OpenAI and Anthropic to enhance complex task capabilities. It supports features such as context-aware auto-replies for WhatsApp, notification summaries, and cross-app operations, and only requires Android 9+ and 8GB of RAM to run.

PokeClaw's Main Features

• Fully Local Offline AI Operation: Performs inference on-device based on the Google Gemma 4 model, requiring no internet access, account, or API key, and data never leaves the phone.
• Optional Cloud Model Enhancement: Supports connecting to OpenAI, Anthropic, Google, and other APIs for stronger reasoning capabilities, displaying token consumption and cost in real time.
• Phone Automation Control: Reads the screen UI tree through Android's accessibility service and autonomously performs actions such as tapping, swiping, entering text, sending messages, and opening apps.
• Context-Aware Auto-Reply: Monitors message notifications from apps such as WhatsApp, automatically opens the chat interface to read the conversation history, and then generates an intelligent reply.
• Skills System (Skills): Provides preset reusable workflows that combine general tools into specific task sequences, such as auto-reply, morning briefing, and smart forwarding.
• System Information Query: Reports battery status, storage space, phone temperature, Bluetooth status, installed apps, Android version, and more in real time.
• Multilingual and Device Compatibility: The interface supports English, Hindi, Japanese, German, and Traditional Chinese, and is compatible with chips such as Tensor G3/G4, Snapdragon 8 Gen 2+, and Dimensity 9200+.

How to Use PokeClaw

• Download and Install: Get the APK file from GitHub Releases to install https://github.com/agents-io/PokeClaw，首次启动时自动下载约 the 2.6GB local model.
• Grant Permissions: Enable the accessibility service permission (required, for screen reading and operations) and notification access permission (recommended, for background message monitoring).
• Choose an Operation Mode: Local mode can be used without any configuration, while cloud mode requires entering an OpenAI, Anthropic, or Google API key in the settings.
• Chat Interaction: Send instructions in natural language using the input box at the bottom.
• Task Cards: In the task interface, tap a skill card (such as "Monitor WhatsApp Messages" or "Send Message"), fill in the contact name and content, and then execute.
• Run in the Background: After enabling auto-reply monitoring, keep the app running in the background, and the AI will automatically detect new messages and execute the reply process.

PokeClaw's Core Advantages

• Fully Local Offline Operation: Based on the Google Gemma 4 model for direct on-device inference, core features can be used without an internet connection, an account, or an API key, and all data processing is completed inside the phone, fundamentally eliminating the risk of privacy leaks.
• Privacy-first architecture: In local mode, the user's screen content, chat history, and personal data are not transmitted to any external server, and sensitive information always remains on the device, achieving true data sovereignty.
• Context-aware intelligence: The auto-reply feature not only monitors message notifications but also actively opens the chat interface to read the full conversation history, generating appropriate replies based on the preceding context rather than simply triggering a fixed template.
• Skills system (Skills): Combines general tools (tap, swipe, input, etc.) into reusable workflows (such as auto-reply and morning briefing), enabling small models to reliably execute complex multi-step tasks while also supporting user-defined skill extensions.
• Transparent and controllable costs: Cloud mode displays token consumption and operating costs in real time, supports setting soft/hard budget limits, and lets users keep track of resource usage at any time during the execution of complex tasks.
• Zero-threshold deployment: No root permission is required, no developer account is required, and it can be used immediately after installing the APK. It supports mainstream devices running Android 9+ with 8GB of memory or more, and is compatible with multiple chip platforms such as Tensor, Snapdragon, and Dimensity.

PokeClaw Key Information and Usage Requirements

• Project nature: An open-source Android AI automation app, licensed under Apache 2.0, independently developed and maintained by developer Nicole, with code hosted on GitHub (agents-io/PokeClaw)
• Core positioning: The world's first phone AI agent to achieve fully local offline operation (On-device AI Agent), capable of autonomously controlling Android devices to perform complex tasks without a cloud connection
• Technical foundation: Based on the Google Gemma 4 model (inferred through the LiteRT-LM runtime), supports native tool calling (Tool Calling), the model size is about 2.6GB, and pure local inference requires no GPU acceleration (can run on CPU)
• Operation modes: A local-first architecture that runs offline by default to protect privacy; optionally connects to cloud APIs such as OpenAI, Anthropic, and Google to enhance complex task processing capabilities, and supports seamless model switching during a session
• Control method: Obtains screen UI tree information through the Android Accessibility Service, and the AI autonomously decides and executes operations such as tapping, swiping, inputting, and sending messages, forming a closed-loop agent process
• Core features: Context-aware auto-reply (supports WhatsApp), notification summaries, clipboard analysis, system information queries, cross-app operations, and a reusable workflow system based on the Skills architecture
• System requirements: Android 9.0 or above, arm64 architecture chip, minimum 8GB RAM (12GB+ recommended), at least 3GB of free storage space (for the first model download), no root permission required
• Hardware performance: Pure CPU devices can run it but startup takes about 45 seconds; devices equipped with chips such as Tensor G3/G4 (Pixel 8/9), Snapdragon 8 Gen 2+, and Dimensity 9200+ are recommended for second-level response
• Required permissions: Accessibility service permission must be granted (for screen reading and automated operations); notification access permission is recommended (for background message monitoring and auto-reply features).
• Initial setup: When local mode is selected for the first time after installing the APK, approximately 2.6GB of model files are automatically downloaded; choosing cloud mode requires configuring the corresponding API key in settings (OpenAI, Anthropic, Google, etc.).
• Language support: The interface supports English, Hindi, Japanese, German, and Traditional Chinese; the skill system supports command input and app operations in any language.
• Network requirements: Local mode runs completely offline without a network; cloud mode requires an internet connection to the corresponding API service; model downloads require a stable network connection (resumable downloads supported).

Analysis of PokeClaw's competitors

PokeClaw's application scenarios

• Automated message management: Monitors WhatsApp and other apps for messages from specific contacts and automatically replies with context-aware responses; supports automatically sending follow-up text messages after missed calls; suitable for scenarios such as driving or meetings where manual replies are not possible, or to help users manage high-frequency social communication.
• Intelligent device status monitoring: Checks battery health, storage space usage, phone temperature, Bluetooth status, and installed apps in real time, automatically analyzes and suggests cleanup targets; suitable for device maintenance, performance optimization, and troubleshooting.
• Cross-app information retrieval and integration: Search for videos within the YouTube app, view trending topics on Twitter/X, find content on Reddit, and copy email subjects for Google searches, suitable for research and information-gathering scenarios that require quickly obtaining information from multiple platforms.
• Personal Schedule and Information Briefing: Execute the morning briefing process, automatically open the weather app to read the temperature, check the calendar to get the day's events, count unread emails, and generate a summary, suitable for information aggregation and itinerary planning at the start of each day.
• Privacy-Sensitive Data Processing: Analyze clipboard content, summarize notification information, and process personal sensitive data in fully offline mode, suitable for business, medical, or legal scenarios with extremely high privacy requirements, ensuring that sensitive information does not leave the device.
• Accessibility Operation Assistance: Enable AI to autonomously complete operations such as tapping, swiping, and typing through natural language instructions, helping users with limited mobility or visual impairments control mobile apps more conveniently and lowering the barrier to smartphone use.