Local model loop
On-device Gemma and MiniCPM-V run inside the phone workflow, with history trimming, cache cleanup, model downloads, and resumable installation built for mobile limits.
Definition
What is a mobile-native agent framework?
A mobile-native agent framework is an agent runtime designed for phone constraints: scoped permissions, tight memory budgets, short sessions, on-device models, and native OS tool calling. Instead of adapting a server-side agent stack to a phone, it treats these limits as first-class design inputs, so agent actions run through the platform's own capabilities and stay inside its permission model.
PhoneClaw is a mobile-native local AI Agent framework for phones and edge devices. Its iOS runtime runs on-device models and native iOS Skills by default, and supports optional Mac Gateway remote inference for heavier local models.
Start from the PhoneClaw overview, then see native mobile Skills for how model decisions become permission-scoped phone actions.
Runtime frame
PhoneClaw starts from the phone runtime.
The framework is shaped by local models, native mobile APIs, short interaction windows, permission prompts, memory budgets, and mobile entry points.
Native Skill loop
Skills turn model decisions into permission-scoped mobile actions across Calendar, Reminders, Contacts, Clipboard, HealthKit, translation, images, and explicit Web Search.
Interaction loop
Text, voice, camera, widgets, Control Center, Dynamic Island, and LiveLand all feed the same task runtime instead of being separate demos.
Edge loop
Mac Gateway extends the runtime over the local network for heavier inference while the phone keeps the agent UI, Skill routing, permissions, and task state.
Core design
Mobile constraints shape the agent runtime.
A mobile agent works with tighter memory, shorter interaction windows, permission prompts, network variation, background limits, and smaller local models. PhoneClaw makes those constraints first-class design inputs.
| Design area | Mobile constraint | PhoneClaw approach |
|---|---|---|
| Runtime | App lifecycle, permission prompts, and phone-scale memory budgets. | iOS app runtime with explicit permissions, lifecycle limits, and local model budgets. |
| Tools | Scoped native APIs and user-granted data access. | Scoped iOS Skills for Calendar, Reminders, Contacts, Clipboard, HealthKit, Translate, Web Search, and images. |
| Interaction | Short sessions, voice, camera, widgets, and task status surfaces. | Text, voice, camera, widgets, Shortcuts, Control Center, Dynamic Island, and LiveLand status. |
| Inference | On-device model size, memory pressure, and optional local edge compute. | On-device Gemma 4 E2B / E4B via LiteRT-LM, MiniCPM-V 4.6, and optional LAN Mac Gateway. |
Framework pieces
PhoneClaw combines model runtime, Skills, and mobile surfaces.
The framework value is in how local inference, native tools, permissions, model state, and mobile entry points work together.
File-driven capabilities
Each Skill describes a focused capability and exposes only the native tools it needs.
Mobile model budget
Model switching, cache cleanup, history trimming, and download handling keep the app usable on device.
Permission boundaries
Personal data sources stay behind iOS permissions and explicit user actions.
Optional Mac Gateway
The phone keeps the agent surface while a trusted Mac can handle heavier local inference.
Developer fit
Where PhoneClaw fits.
PhoneClaw is useful when a project needs the phone itself to host the agent surface, local model path, native capability layer, and optional local-network inference.
Best fit
Users who want a mobile-native local agent framework for phones and edge devices, with native mobile Skills, local model execution, and privacy-specific data boundaries.
Operating model
PhoneClaw works through iOS permissions, native APIs, local models, and focused mobile workflows.
Developer angle
Open-source developers can inspect the Swift app, Skill files, model runtime choices, and Mac Gateway.