kellyvv/PhoneClaw

What it solves

PhoneClaw transforms a smartphone into a local AI agent runtime, allowing users to interact with their personal data and device functions using natural language while keeping data processing on-device for privacy.

How it works

It runs lightweight LLMs (like Gemma 4 E2B/E4B) and multimodal models (MiniCPM-V 4.6) directly on iOS using LiteRT. The system uses a file-driven "Skill" system where capabilities—such as accessing the calendar, reminders, and health data—are defined in Markdown files. It also supports an optional Mac Gateway for remote inference via Ollama or other CLI tools over a local network.

Who it’s for

iOS users with modern hardware (A16/A17 Pro chips) who want a private, offline-first AI assistant capable of managing device tasks and analyzing personal health and schedule data.

Highlights

• On-Device Execution: Runs inference and skill calls locally to ensure privacy and eliminate per-call token costs.
• Native iOS Integration: Built-in skills for Calendar, Reminders, Contacts, Clipboard, and HealthKit data.
• Multimodal Capabilities: Supports image Q&A and a "LIVE" mode for real-time camera recognition and voice interaction.
• Flexible Inference: Offers a choice between fully local execution or remote inference via a paired Mac on the same LAN.
• LiveLand: Integration with the Dynamic Island and widgets for quick access to the AI agent.