🦖 tinyArms

Native Apple ecosystem AI assistant. Runs on-device, costs nothing, learns your preferences.

Super lightweight. Designed to be invisible. Learns from you. Grows with you. Task-specific. Makes your life easier without replacing your brain.

Platforms: macOS (menu bar daemon) • iOS (Share Sheet) • iPadOS (Split View)

---

⚠️ Status: Swift Migration (Phase 1)

What's happening: Full rewrite from TypeScript → Swift for native Apple ecosystem integration.

What works: Architecture design, research complete, Swift foundation started.

What's coming:

• Phase 1: macOS daemon (LaunchAgent, FSEvents, Core ML)
• Phase 2: iOS Share Extension (screenshot renaming, OCR, privacy redaction)
• Phase 3: iPadOS enhancements (Split View, Drag & Drop)
• Phase 4: CloudKit sync (results across devices)

Timeline: Phase 1 (Q1 2026), iOS beta (Q2 2026), App Store (Q3 2026)

⭐ Star this repo to follow Swift development.

---

The Vision

Imagine you take a screenshot on iPhone. tinyArms reads it via Share Sheet, extracts text via OCR, and offers 3 intelligent names:

Screenshot_2025-11-02.png

tinyArms suggests:
  1. hero-section-mobile-v1.png
  2. landing-page-screenshot.png
  3. website-mockup-draft.png

You pick #1 (or type your own).

tinyArms learns: "This user prefers structured names with platform + version."

Next screenshot → Better suggestions. Every choice trains the system.

Cross-device sync: Rename on iPhone, search from Mac via Spotlight (CloudKit sync).

---

What Makes 🦖 tinyArms Different

1. Native Apple Ecosystem Integration

macOS: Menu bar daemon, Spotlight indexing, Quick Look previews iOS: Share Extension, Shortcuts/Siri, Widgets, Live Activities iPadOS: Split View, Drag & Drop batch processing Sync: CloudKit (results searchable across all devices)

2. 100% On-Device (Privacy-First, Zero Cost)

• macOS: Local LLM inference (Ollama or MLX Swift)
• iOS/iPadOS: Core ML models (SmolLM2-360M, MobileBERT, CLIP)
• No cloud: Your code, files, data never leave your devices
• No API keys: No OpenAI/Anthropic accounts needed
• No subscriptions: Free tier forever (Pro tier optional for advanced skills)

Model stack:

• macOS: Qwen2.5-Coder-3B via Ollama or MLX Swift (1.9GB)
• iOS: Core ML SmolLM2-360M (250MB, Apple Neural Engine optimized)

3. Tiered Routing (Fast Rules + Smart AI)

Not everything needs AI. 🦖 tinyArms routes tasks intelligently:

Level 0: Deterministic Rules (<1ms, 60-75% of tasks)
  - Hardcoded color detection: bg-[#3B82F6]
  - File type detection, kebab-case formatting
  ↓ (no match)

Level 1: Semantic Routing (<100ms, 20-25% of tasks)
  - Core ML embeddings (MobileBERT 100MB)
  - Intent classification, similarity search
  ↓ (complex task)

Level 2: On-Device LLM (2-3s, 10-15% of tasks)
  - macOS: Qwen2.5-Coder-3B (code linting, refactoring)
  - iOS: SmolLM2-360M (grammar, simple rewrites)
  ↓ (deep analysis)

Level 3: Optional Large Model (10-15s, <5% of tasks)
  - macOS: Qwen2.5-Coder-7B (architectural violations)
  - iOS: Cloud fallback (optional, explicit user consent)

Result: Simple tasks instant, complex tasks 2-3s. Best of both worlds.

4. Self-Improving via Prompt Evolution (Future)

Planned for Phase 5: Adaptive prompts that improve through user feedback.

• Accuracy drops? System generates 3 new prompt variants (offline)
• You vote on outputs (Thompson Sampling A/B test)
• Best prompt auto-promoted
• 99% offline, <$0.06/year cloud cost

Reference: ideas/future-prompt-evolution.md

---

Hardware Requirements

Platform	RAM	Storage	Performance
macOS	8GB (16GB recommended)	2-6GB models	M1+ (Apple Silicon required)
iOS	4GB+	500MB-1GB models	iPhone 12+ (A14+ for Neural Engine)
iPadOS	6GB+	500MB-1GB models	iPad Pro 2020+ or iPad Air 4+

Platforms: macOS 13.0+ (Ventura), iOS 17.0+, iPadOS 17.0+

---

Quick Start (When Released)

macOS

# Download from GitHub Releases
open TinyArms-macOS-v0.2.0.dmg

# Grant permissions (LaunchAgent, File Access, Spotlight)
# Menu bar icon appears → Ready

# Or build from source:
git clone https://github.com/nqh/tinyArms
cd tinyArms
open TinyArms.xcodeproj
# Build & Run (Cmd+R)

iOS/iPadOS

1. Install TestFlight from App Store
2. Scan QR code (beta invite)
3. Open tinyArms app
4. Take screenshot → Share → tinyArms
5. Rename suggestion appears → Tap "Apply"

---

Core Skills

1. code-linting (macOS)

   • Pre-commit hook integration
   • Detects: hardcoded colors, magic numbers, file size >350 LOC
   • Speed: 2-3s per file (Qwen2.5-Coder-3B)

2. privacy-redaction (iOS/macOS)

   • Auto-detect PII (emails, phone numbers, API keys)
   • Blur before sharing (Share Extension integration)
   • Pre-commit hook (block commits with secrets)

3. context-aware-clipboard (macOS)

   • Semantic search clipboard history (MobileBERT embeddings)
   • Auto-format on paste (URL → markdown, JSON → pretty-print)

4. file-naming (iOS/macOS)

   • Screenshot → descriptive names
   • Learns from your choices (prompt evolution)

---

Documentation

Foundation

• 00-SWIFT-QUICKSTART.md - 5-minute Xcode setup
• 01-SWIFT-ARCHITECTURE.md - TinyArmsKit package structure

Platform Guides

• 02-MACOS-DAEMON.md - LaunchAgent, FSEvents, menu bar
• 03-IOS-PLATFORM.md - Share Extension, Shortcuts, Widgets
• 04-IPADOS-PLATFORM.md - Split View, Drag & Drop

Advanced

• 05-COREML-MODELS.md - Model conversion, Neural Engine
• 06-CLOUDKIT-SYNC.md - Cross-device sync architecture
• 08-APP-STORE-DEPLOYMENT.md - Code signing, TestFlight

Research

• research/ - Tiered routing, confidence scoring, semantic caching (11 papers)
• research/06-apple-foundation-models-integration.md - Apple FM vs tinyArms analysis

---

Development Status

Phase 1: macOS Foundation (In Progress)

• Swift architecture design
• Research complete (tiered routing, industry validation)
• Xcode project setup (TinyArmsKit package)
• LaunchAgent daemon
• FSEvents file watching
• Ollama/MLX Swift integration
• Menu bar app (SwiftUI)

Phase 2: iOS MVP (Planned Q2 2026)

• Share Extension (image, text, file input)
• Core ML models (SmolLM2-360M, MobileBERT)
• Result UI (SwiftUI card)
• CloudKit sync (basic)

Phase 3: iPadOS + Advanced (Planned Q3 2026)

• Split View support
• Drag & Drop batch processing
• Shortcuts integration (Siri)
• Widgets + Live Activities

Phase 4: App Store Release (Planned Q4 2026)

• Code signing + notarization
• TestFlight public beta
• App Store submission
• StoreKit 2 (subscriptions)

---

Why "tinyArms"?

Like a T-Rex: Small arms, but surprisingly capable when working WITH larger tools.

tinyArms doesn't replace your brain or workflow. It's the invisible AI assistant that learns your preferences, handles repetitive tasks, and gets better over time—without you thinking about it.

Platform-first: Designed for how you already work (Share Sheet, Shortcuts, Spotlight, menu bar).

---

Contributing

See CONTRIBUTING.md for Swift development setup.

Requirements:

• Xcode 15.0+ (macOS Sequoia SDK)
• Apple Developer account ($99/year for code signing)
• Swift 5.9+
• macOS 14.0+ (development machine)

---

License

MIT 🦖

Native Apple ecosystem. On-device ML. Learns from YOU.