swift-agents

Swift Agents

16 specialized Swift agents for Claude Code and VS Code Copilot.

Built by Taylor Arndt at Techopolis. Swift 6.2 strict concurrency, Apple Foundation Models, on-device AI, SwiftUI best practices, and mobile accessibility — enforced on every prompt.

Agents handle the review workflow. Marketplace skills provide the deep reference knowledge. This keeps agents slim, fast, and focused.

The Problem

AI coding tools write Swift like it is 2020. They use ObservableObject when @Observable exists. They ignore actor isolation. They produce views with no accessibility modifiers. They have never heard of Apple Foundation Models or @Generable.

Before and After

Without Swift Agents:

class SettingsViewModel: ObservableObject {
    @Published var notifications = false
}

struct SettingsView: View {
    @StateObject var vm = SettingsViewModel()
    var body: some View {
        NavigationView {
            Toggle("Notifications", isOn: $vm.notifications)
        }
    }
}

With Swift Agents:

@Observable
class SettingsViewModel {
    var notifications = false
}

struct SettingsView: View {
    @State private var vm = SettingsViewModel()
    var body: some View {
        NavigationStack {
            Toggle("Notifications", isOn: $vm.notifications)
                .accessibilityLabel("Push notifications")
                .accessibilityHint("Enables or disables push notifications for this app")
        }
    }
}

Modern @Observable instead of ObservableObject. NavigationStack instead of deprecated NavigationView. Accessibility labels and hints so VoiceOver users know what the toggle does.

The Team

Agent	Role
swift-lead	Orchestrator. Routes tasks to the right specialists.
concurrency-specialist	Swift 6.2 strict concurrency, actors, Sendable, async/await.
foundation-models-specialist	Apple Foundation Models (iOS 26+), @Generable, @Guide, tool calling.
on-device-ai-architect	MLX Swift, llama.cpp, Core ML, device tier planning, multi-backend fallback.
mobile-a11y-specialist	iOS/macOS accessibility, VoiceOver, Dynamic Type, focus management.
swiftui-specialist	Modern SwiftUI, @Observable, state management, NavigationStack.
app-review-guardian	App Store review compliance, privacy manifests, IAP rules.
testing-specialist	Swift Testing, XCTest, testable architecture.
swift-security-specialist	Keychain, CryptoKit, biometric auth, ATS, certificate pinning.
coreml-specialist	Core ML conversion, quantization, deployment.
mlx-specialist	MLX Swift, MLXLLM, MLXVLM, GPU memory, LoRA.
meta-glasses-sdk-specialist	Meta Wearables DAT, camera streaming, device pairing.
swiftdata-specialist	SwiftData persistence, @Model, @Query, migration.
visionos-specialist	visionOS, RealityKit, ARKit, immersive spaces.
storekit-specialist	StoreKit 2, subscriptions, entitlement verification.
performance-specialist	Instruments, MetricKit, hang detection, memory management.

How It Works

A UserPromptSubmit hook fires on every prompt. If the task involves Swift code, the hook delegates to the swift-lead. The lead evaluates the task and invokes the relevant specialists. Multiple specialists can review a single task.

Agents are slim workflow reviewers (~50-80 lines each). They flag issues and enforce patterns. For deep reference knowledge, they rely on marketplace skills that load on demand. The installer offers to install companion skills automatically.

Prerequisites

For Claude Code:

• Claude Code CLI installed
• A Claude Code subscription (Pro, Max, or Team)

For VS Code Copilot:

• VS Code (1.106+) or VS Code Insiders
• GitHub Copilot extension
• A GitHub Copilot subscription

Installation

One-Liner (Recommended)

curl -fsSL https://raw.githubusercontent.com/Techopolis/swift-agents/main/install.sh | bash

The installer asks which tool(s) and where (project or global). It also offers to install companion skills from the marketplace.

Non-Interactive