Edge Scaffold

Edge Scaffold generates a ready-to-build reference agent from an optimized model. It shows the recommended iOS integration for Edge Kit, Edge Halo, EdgeMesh, EdgeData, and Neural Imprint.

Developer Preview

Generated apps require signing, device testing, and store review before release.

How it works

Edge Studio (export) → Edge Scaffold (template + config) → Xcode project → App

Edge Studio writes a ZIP containing the app template, model metadata, runtime configuration, and optional personalization resources. You unzip, configure signing, build, test on a real device, and then own the resulting app project.

ScaffoldConfig.swift

All app behavior is controlled by one file:

enum ScaffoldConfig {
    static let appName = "MyApp"
    static let appDescription = "A private on-device assistant."
    static let defaultSystemPrompt = "You are a concise assistant."
    static let modelCategory: ModelCategory = .llm  // .llm | .vlm | .tts
    static let bundleModelName: String? = "Qwen3.5-9B-4bit"
    static let defaultTTSSpeaker: String? = nil
}

Field	Controls
appName	Display name in app UI
modelCategory	Which engine and UI path the app uses
bundleModelName	Bundle folder name when shipping model inside the app
defaultSystemPrompt	Initial system instruction for chat

The app UI automatically adapts based on modelCategory:

Category	Input	Output
LLM	Text	Streaming text
VLM	Text + photo	Streaming text
TTS	Text	Audio

Reference personalization lanes

Edge Scaffold includes developer-facing settings screens for the full local personalization lifecycle:

Lane	What it demonstrates
Base Model	Load model, inspect runtime state, toggle Neural Imprint restore for A/B checks.
Tool Protocol Learning	Register app-owned read-only tools and test tool-call behavior without putting live tool schema into every prompt.
User Profile Learning	Run the profile workflow from local sample data and activate a combined Neural Imprint artifact.
Correction Learning	Capture feedback and corrections as local data that can feed later profile refreshes.
EdgeMesh	Pair with a Mac, upload receipts, receive compatible capsules, and report apply status.

Use these lanes as reference code. Replace the sample data and sample tools with your agent's own domain model.

Build and test

1. Unzip the export. Open in Xcode.
2. Select your development team and a unique bundle identifier.
3. Set a real device as run destination — not simulator.
4. Build as Release for performance validation.
5. Test: first launch, model load, first response, multi-turn, backgrounding.

For larger models, enable the Increased Memory Limit entitlement and validate with a Release build on the target device.

Model delivery

Tier	When to use
Bundle	Small models (< 2GB). Ships inside the app binary.
On-Demand Resources	Medium models. Downloaded after install, managed by the OS.
HuggingFace	Large models. Downloaded on first launch from HuggingFace.
Cache	Previously downloaded models. Fastest path on subsequent launches.

What to customize

Start with:

1. ScaffoldConfig.swift for product name, model ID, model category, sample domain, and bundled resources.
2. The sample data provider for your agent's local facts.
3. The tool registry for your app-owned read-only tools.
4. The settings copy that explains personalization and reset behavior to your users.

Do not copy dogfood-specific business logic into a production agent. Keep app policy in the app layer and use Edge Kit / Edge Halo only for reusable infrastructure.

Next steps

• Build and ship example — End-to-end walkthrough from optimize to App Store.
• Platform requirements — Device and OS constraints.