GitHub - Saladino93/hitokudraft: Hitoku Draft. A context aware macOS local AI assistant. Fully local.

A context aware native macOS menu bar AI assistant.

Everything runs locally on Apple Silicon. No cloud calls.

Hold to talk. Ask it for email drafting, or understanding a PDF, or set up a calendar meeting.

Download here https://hitoku.me/draft/ or compile yourself.

⚠️ Gemma 4 (LiteRT) — Memory Warning

Gemma 4 models run via Google's LiteRT engine with a WebGPU backend. On Apple Silicon, the WebGPU runtime can allocate significantly more GPU memory than the model weights alone — we've observed up to 10x the model size in Metal buffer allocations during inference. On machines with 16 GB or less, this may cause system instability. We recommend using Gemma 4 E2B (smaller) over E4B, keeping prompts short, and monitoring Activity Monitor. There is also a known upstream crash in the WebGPU deallocation path (dawn::SlabAllocatorImpl) that is inside Google's prebuilt dylibs and cannot be fixed downstream. See LiteRT issue #5706. We are actively working on migrating Gemma 4 to MLX for stable, predictable memory usage.

Features

• Context Awareness — Reads surrounding screen context (active app, selected text, document content) to produce better results.
• Draft Mode — Generate new content from voice ("draft an email declining the meeting") — pastes at cursor.
• Voice Edit — Select text, press a hotkey, speak an instruction ("make this more formal", "translate to Spanish"), and the text is rewritten in-place.
• Grammar Fix — One-hotkey grammar and spelling correction with automatic language detection (English, Italian, French, German, and more).
• Continuous Dictation — Real-time speech-to-text with optional LLM polish (filler removal, punctuation).
• Action Mode — Voice-triggered actions: set timers, create calendar events, send emails, take notes.
• Web Search & Tools — LLM can search the web, fetch URLs, and query your calendar to answer questions.
• Multiple STT Backends — Parakeet TDT (CoreML), Qwen3-ASR (streaming MLX), WhisperKit (Whisper tiny/base).
• Multiple LLM Backends — MLX (Qwen3.5), LiteRT (Gemma 4 with native audio + vision).
• Text-to-Speech — Reads results aloud via Kokoro or PocketTTS when the focused element isn't editable.
• Display Mode — Shows results in a floating overlay with syntax highlighting, LaTeX math rendering, and code blocks when you're not in a text field.

Requirements

• macOS 14.0+
• Apple Silicon (M1 or later)
• Xcode 16.0+ (to build from source)
• Microphone and Accessibility permissions

Building from Source

The project uses xcodegen to manage the Xcode project:

# Install xcodegen if you don't have it
brew install xcodegen

# Generate the Xcode project
xcodegen generate

# Build
xcodebuild -scheme HitokuDraft -configuration Release build

LiteRT dylibs for Gemma 4 support are not included in the repository. To enable LiteRT:

cd HitokuInference
./setup_litert_libs.sh

Architecture

• ConversationCoordinator — Single orchestration center (state machine + pipeline sequencing)
• Protocol-backed services — STTService, LLMService, TTSProvider, EditabilityDetector
• HitokuInference — Unified inference routing across MLX and LiteRT backends
• AppState enum — Single source of truth for all UI state
• Swift concurrency — async/await and actors throughout; @MainActor on coordinators

See docs/SWIFT_REWRITE_BRIEF.md for the full architecture specification.

Models

All models are downloaded from HuggingFace on first use — nothing is bundled in the app binary.

Acknowledgments

See ACKNOWLEDGMENTS.md for the full list of open-source libraries, AI models, and their licenses.

License

See LICENSE for details.