projects/ideai

ideAI

A mirror with memory

Your phone is the most trusted device you own. ideAI treats it that way — all intelligence runs on-device. You journal, it reflects. It finds patterns in your thinking you didn't know were there. Auto-tagging, semantic search, multi-note synthesis. Military-grade encryption, zero cloud dependency.

Visit ideaapp.com (coming soon) App Store

Why on-device

The cyberpunk inversion already happened. Your phone's Secure Enclave is the most trusted computing environment you own — more than your laptop, infinitely more than any cloud. ideAI builds on that foundation. Every model runs locally via MLX. Your journal entries never leave the device. The AI that understands your thinking patterns is yours alone.

What it does

  • Semantic search across all your notes — under 500ms, entirely on-device
  • Auto-tagging via embeddings — under 100ms per note
  • Image attachments with on-device vision tagging — photos become searchable alongside text
  • Multi-note synthesis: the AI reads 10-20 of your notes and surfaces patterns you missed
  • AES-256-GCM encryption with biometric unlock (Face ID / Touch ID)
  • MCP server integration — Claude can read your journal with your permission

In action

Data Shard Mode — MCP server integration with Claude.

1 / 3

Architecture

16 Swift packages organized by feature. Each module builds and tests independently. The app coordinates three inference tiers — routing between them automatically based on availability, device capability, and task priority.

features: Notes, Search, Synthesis, Atlas (graph). InferenceCoordinator (priority queue + preemption). inference tiers: Foundation Models (Apple Intelligence), MLX Swift (19 models, 350M to 8B), Cloud relay (Claude via Cloudflare Workers). storage: SwiftData (iCloud store + local store), AES-256-GCM (Keychain, Face ID / Touch ID)

Three inference tiers with priority-based coordination. User-facing ops always preempt background tasks.

Key decisions

  1. Priority-based inference coordination

    conversation preempts compression, preventing OOM on a phone with a single GPU

  2. Dual-store SwiftData

    user content syncs to iCloud, regenerable data (embeddings, caches) stays local

  3. Native Apple embeddings first, MLX fallback

    zero app-size cost for semantic search on supported devices

  4. Swift 6 strict concurrency across all 16 modules

    data races caught at compile time, not in production

  5. MCP server built in

    Claude can read your journal with your permission, over an encrypted WebSocket tunnel