Hitokudraft Puts a Privacy-First Voice AI Right in Your Mac Menu Bar
A new open-source project called Hitokudraft brings a fully local, voice-first AI assistant directly to the macOS menu bar. This native application runs entirely on Apple Silicon, processing voice commands and screen context without any data leaving your computer. It lets you draft emails, edit text, understand PDFs, and manage your calendar using just your voice.
Developer Saladino93 designed the tool to handle everyday productivity tasks by understanding what you are doing on screen. The assistant reads your active application, selected text, and document content to give more accurate responses. Everything from speech recognition to text generation happens locally using a mix of MLX and LiteRT backends.
Key capabilities of the assistant
- Voiced draft generation that pastes at cursor.
- In-place voice editing for selected text.
- One-hotkey grammar fixes with language detection.
- Continuous real-time dictation with optional polish.
- Voice-triggered actions for calendars and emails.
- Web search and URL fetching capabilities.
- Floating overlay with syntax highlighting and math.
This tool is built for privacy-conscious professionals who want a capable AI assistant without sending data to cloud servers. Small agency workers using local AI for daily tasks will find the context-aware screen reading and voice editing helpful for drafting and revising content. Prosumers with Apple Silicon Macs can also benefit from direct model choices, picking from several speech recognition and text generation backends based on their hardware and task.
Memory considerations and next steps
The developer has flagged a significant memory warning for users wanting to run Gemma 4 models through the LiteRT backend. On Apple Silicon, the WebGPU runtime can allocate up to ten times the model size in GPU memory during processing, which may cause system instability on machines with 16 GB of RAM or less. To address this, the creator recommends using the smaller Gemma 4 E2B model and is actively working on migrating Gemma 4 support to the MLX framework for more stable and predictable memory usage.
"It's context-aware; it reads your screen, documents, and active app to understand what you're working on." — Source: Reddit