PokeClaw, an open-source Android agent running entirely on-device, launched on GitHub on April 6, 2026, gaining 253 stars. Short for PocketClaw, the project represents a shift toward privacy-first mobile AI automation by processing all inference locally without cloud dependencies or API keys.
Gemma 4 Powers Local Phone Control Through Accessibility APIs
PokeClaw uses Gemma 4 running locally via LiteRT (Google's on-device inference runtime) combined with Android's Accessibility Service to observe and interact with phone UI elements. The architecture enables the agent to "see" screen content through UI tree analysis and execute actions like tapping, swiping, and typing based on user requests.
Technical specifications include:
- Powered by Gemma 4 with all inference happening on-device
- Uses Android Accessibility Service for UI observation and interaction
- Supports tool calling for app control
- No cloud dependency or API keys required
- Written in Kotlin
- Works completely offline
Privacy-First Design Eliminates External Data Transmission
Unlike cloud-based phone assistants such as Google Assistant or Siri, PokeClaw processes everything locally, eliminating privacy concerns about sensitive information being uploaded to external servers. All user data, commands, and interactions remain on the device, giving users full control over their information.
The local-first architecture offers several advantages:
- No data sent to external servers
- Full offline functionality
- No privacy risks from cloud processing
- Complete user control over the AI model
- No dependency on third-party services
Simultaneous iOS Project Suggests Growing Interest in Edge AI Automation
A parallel project called PhoneClaw by developer kellyvv emerged around the same time, accumulating 358 GitHub stars. Written in Swift for iOS and also using Gemma 4 for on-device phone control, the simultaneous emergence of these projects indicates growing interest in privacy-preserving mobile AI agents.
The timing reflects broader industry momentum toward edge AI automation. With Gemma 4's 2B parameter model being small enough to run efficiently on mobile devices while remaining capable enough for agent tasks, developers can now build AI automation tools that don't require sending user data to cloud servers.
Agents-IO Organization Signals Broader On-Device Agent Movement
The agents-io GitHub organization behind PokeClaw appears to be developing multiple on-device agent projects, suggesting this is part of a larger movement toward privacy-first AI automation. The project's GitHub topics include ai-agent, android-automation, gemma4, litert, local-llm, on-device-ai, open-source, phone-agent, pocketclaw, pokeclaw, and tool-calling.
This development marks a significant shift from the cloud-dependent AI assistant paradigm that has dominated mobile platforms, offering users an alternative that prioritizes privacy and local control over their AI-powered automation tools.
Key Takeaways
- PokeClaw launched on April 6, 2026, as an open-source Android agent running entirely on-device with Gemma 4, gaining 253 GitHub stars
- The agent uses Android Accessibility Service combined with local LLM inference to observe and control phone UI without cloud dependencies
- All processing happens locally, eliminating privacy concerns about sensitive data being sent to external servers and enabling full offline functionality
- A parallel iOS project called PhoneClaw with 358 stars emerged simultaneously, indicating growing interest in privacy-first mobile AI agents
- The agents-io organization behind PokeClaw appears to be building multiple on-device agent projects as part of a broader edge AI automation movement