A Private AI Workspace
Mawj (موج) is built for work that should not leave your device. Conversations, documents, images, microphone recordings, and transcripts stay local for AI processing.
Download a model once, then keep working offline. Network access is limited to actions you choose, such as downloading a public model or importing a web link you shared.
- 1Private conversations: Chat history and local preferences are stored on your device.
- 2Private inputs: Documents, images, and dictated speech are processed locally.
- 3Your control: Export or delete your workspace and choose exactly which models to download.
Work with your own context
Import PDFs, Word documents, text, Markdown, and RTF files, then attach them to a conversation. Mawj processes and retrieves relevant passages locally so you can ask focused questions about your material.
Voice input is transcribed locally with WhisperKit, and the share extension can bring selected files and links into the workspace. Vision-capable models can also work with images in a conversation.
The result is a practical, offline-capable workspace for reviewing sensitive documents, research, writing, and everyday thinking.
Architecture: local models, managed responsibly
Mawj manages local inference around the device rather than treating the phone as a thin client.
Multi-engine inference
The MultiEngineRouter uses MLX for local models and can use Apple Foundation Models on supported devices.
• Local models: → MLX on Apple Silicon• Apple models: → Foundation Models when available• User choice: → model preferences are respected
A model library you control
Browse supported model families, download models in the background, and select the right local capability for the task. Mawj keeps model files on-device and makes their status visible.
Device-aware lifecycle
Built with SwiftUI and SwiftData, with explicit memory-pressure, thermal, and background lifecycle handling for local inference.
The workspace in action
