Mobile-native local AI agent framework

PhoneClaw for iPhone

An iPhone Agent built for real mobile constraints: on-device models, native iOS Skills, scoped permissions, Live interaction, and optional Mac inference when the phone should stay light.

This is PhoneClaw for iPhone, the iOS on-device agent project.

Product

A phone-first agent surface.

PhoneClaw treats the iPhone as the runtime for local inference, native Skills, permission checks, model downloads, memory budgeting, and Live interaction modes.

Local-first by default

Gemma 4 E2B / E4B via LiteRT and MiniCPM-V 4.6 run on-device for private everyday tasks.

Native iOS Skills

Calendar, Reminders, Contacts, Clipboard, HealthKit, translation, images, and explicit Web Search.

Live and LiveLand

Voice, camera-aware LIVE mode, and Dynamic Island status for mobile tasks in motion.

Mobile runtime work

Memory budgeting, model switching, resumable downloads, cache cleanup, and history trimming.

Mac as an edge node

Pair a Mac over LAN to use Ollama or CLI models while the iPhone keeps the agent experience.

Framework

The agent stack is shaped by iOS limits.

Desktop agent frameworks assume long-running processes, generous memory, broad file access, and easy background execution. PhoneClaw starts from the opposite constraints.

01 Input

Mobile commands

Text, voice, images, LIVE camera, widgets, Shortcuts, Control Center, and LiveLand entry points.

02 Routing

Skill-aware planning

File-driven Skills, scoped tool lists, argument extraction, clarification, and safer multi-turn task handling.

03 Execution

Native iOS tools

Calendar, Reminders, Contacts, Clipboard, HealthKit, Web Search, and device-local summaries behind permission gates.

04 Inference

On-device or LAN

LiteRT local models by default, with optional Mac Gateway for heavier models on a trusted local machine.

Boundaries

The privacy claim is concrete.

PhoneClaw keeps chat, images, and personal data on iPhone by default. Explicit user actions add network paths: Web Search, opening a URL, or Mac remote inference.

Stays on iPhone

Default inference, native Skill calls, Calendar, Reminders, Contacts, Clipboard, HealthKit summaries, images, and chat.

Leaves only when asked

Realtime Web Search and opening a public URL send the request to the search provider or target website.

Optional Mac inference

When paired, the request goes to your Mac on the LAN. Ollama stays there; CLI providers follow their own policies.

Read the full Privacy & Data Flow page

Mac Gateway

Keep the phone native. Move heavy inference to your Mac.

PhoneClaw Gateway turns a Mac on the same Wi-Fi into an optional local inference source. The iPhone still owns the chat UI, Skills, permission model, and task flow.

Deep dives

Technical proof for the main claims.

These pages are written for developers, search, and answer engines that need specific evidence instead of a product tagline.

Mobile-native Agent framework

Why PhoneClaw is designed from iOS permissions, memory, mobile entry points, and edge inference rather than copied from desktop Agent stacks.

Gemma 4 on iPhone

LiteRT-LM, MiniCPM-V 4.6, mobile memory limits, and how model choice shapes local agent workflows.

Native iOS Skills

File-driven Skills, tool allowlists, permission boundaries, and local tool calling on iPhone.

LiveLand and LIVE mode

Dynamic Island, widgets, voice, camera, and mobile task status for an iPhone-native agent.

PhoneClaw FAQ

Direct answer-engine facts for entity identity, privacy, mobile agent architecture, models, Skills, and Mac Gateway.

Languages

Localized entry pages for Chinese and Japanese search.

PhoneClaw already has zh/en/ja product localization. The website now exposes canonical Chinese and Japanese landing pages with hreflang, so search engines can match language intent directly.

中文入口

面向“iPhone 本地 AI Agent”“移动端 Agent 框架”“端侧 AI 助手”等中文查询。

日本語ページ

Targets Japanese searches for local iPhone AI agents, on-device AI, and private mobile assistants.

llms.txt

A compact machine-readable summary for answer engines and AI crawlers that inspect project context files.