Summary

PhoneClaw provides a fully on-device, private AI agent for iPhones, leveraging Gemma 4 for local inference without cloud reliance. It targets privacy-conscious iPhone users seeking AI assistance, offering benefits like offline operation, enhanced data security, and customizable capabilities through a unique file-based skill system.

How It Works

This project implements a local AI agent using Gemma 4 (E2B/E4B variants) via the MLX framework within a native Swift iOS application. Its core innovation lies in a dynamic, file-based "Skill" system, where AI capabilities are defined in Markdown (SKILL.md) files, allowing users to add or modify features without recompiling the app. The architecture prioritizes privacy, ensuring all data processing, including multi-modal inputs like images, remains strictly on the device.

Quick Start & Requirements

• Environment: macOS, Xcode 16, iOS 17+, CocoaPods. Requires an Apple Developer account for signing on physical devices.
• Installation:
  1. IPA: Download release IPA from GitHub Releases, sign, and install. Models are downloaded on-device post-installation.
  2. Source: Clone repository, run pod install, and build via Xcode.
• Model Management: Gemma 4 E2B (~3.58 GB) and E4B (~5.22 GB) models can be downloaded directly within the app or pre-bundled during the build process.
• Links: GitHub Issues for bug reports and feature suggestions.

Highlighted Details

• On-Device Multi-modal Understanding: Processes photos for recognition, chart interpretation, and scene description locally.
• Dynamic Skill System: Extensible capabilities via SKILL.md files, supporting both device API calls and pure prompt-based tasks.
• Complete Offline & Privacy: Guarantees no data leaves the device, ensuring user privacy.
• Flexible Model Handling: Supports multiple Gemma 4 variants with on-device download or bundling options.
• Recent Additions: Includes voice input/output (April 7, 2026), a "thinking mode," and session history management.

Maintenance & Community

The project shows active development with frequent updates in April 2026, indicating ongoing refinement. Community interaction is primarily channeled through GitHub Issues for bug reports and feature requests.

Licensing & Compatibility

Licensed under the Apache 2.0 license, which generally permits commercial use and integration into closed-source projects.

Limitations & Caveats

The E4B model variant is recommended for iPhone 15 Pro and newer devices. Multi-turn tool calling functionality is still under optimization. Building from source requires macOS and Xcode, and deploying to a device necessitates an Apple Developer account and signing. Model files are substantial, requiring significant device storage.