Your iPhone, but it actually knows your calendar

What it does

PhoneClaw is an on-device iOS agent built around Google’s Gemma 4 and MiniCPM-V models. It handles calendar queries, reminder creation, contact lookups, clipboard operations, health data summaries, and even real-time web searches — all through natural language. The catch: everything runs locally on your iPhone, with no API keys, no cloud inference, and no data leaving the device.

The interesting bit

The project treats “Skills” as plain Markdown files. Want the agent to do something new? Write a SKILL.md, drop it in, and the app picks it up without recompilation. That’s a genuinely modular approach to mobile agent capabilities, not just another chat wrapper. The LIVE mode adds real-time voice conversation with optional camera input, so the model can literally see what you see and talk back.

Key highlights

• Fully offline inference using LiteRT-LM; Gemma 4 E2B (~2.4 GB) for lighter tasks, E4B (~3.4 GB) for multi-turn agent work
• MiniCPM-V 4.6 handles image Q&A and live camera recognition in LIVE mode
• HealthKit integration reads steps, sleep, heart rate, and workouts locally — the model chooses which of 9 health tools to invoke based on your phrasing
• Cross-round KV cache reuse cuts response latency on follow-up questions by roughly 3.5×
• Memory budgeting dynamically adjusts to available RAM, with explicit warnings that sideloaded builds face stricter limits (E4B may require CPU-only inference)
• Bilingual UI and skill descriptions with automatic language switching

Caveats

• iPhone 15 Pro or better recommended for E4B and LIVE camera modes; A16 is the practical floor for E2B
• Sideloadly-signed installs hit memory walls that Xcode builds don’t, forcing CPU fallback or smaller models
• MTP speculative decoding is marked experimental and only benefits short Gemma 4 E4B replies

Verdict

Worth a look if you want a private, offline Siri alternative that actually understands your personal data. Skip it if you’re on older hardware or expect the polish of a cloud-backed assistant — this is a research-grade project with rough edges around memory management and model deployment.