Yooz Engine

Sovereign Intelligence. Built for the skeptical.

Yooz Engine is a unified local AI service for macOS. It runs as a menu bar app on localhost:19920 and provides on-device speech-to-text, LLM inference, grammar correction, voice activity detection, engine-hosted long-context sessions, and (soon) text-to-speech to every Yooz app on your system. Models stay on your device; nothing is sent to the cloud.

This is the source repository. Built with Swift 5.9 (with Swift 6 concurrency idioms), SwiftUI, Hummingbird for HTTP/WebSocket, and MLX-Swift for on-device inference on Apple Silicon.

What's inside

YoozEngine.app — menu bar service exposing localhost:19920
├── /v1/health, /v1/models, /v1/modules     — service introspection
├── /v1/stt/{languages,status,load,batch}   — speech to text (REST)
├── /v1/stt/stream                          — speech to text (WebSocket)
├── /v1/llm/generate                        — LLM text generation
├── /v1/infinite/{models,status,sessions}   — long-context sessions
├── /v1/touchup                             — STT cleanup pipeline
├── /v1/grammar/check                       — rule-based grammar
└── /v1/vad/detect                          — voice activity detection

Build variants pick which modules ship (YoozEngine full, YoozEngineWhisper no VAD or Infinite, YoozEngineLite no MLX STT, VAD, or Infinite). All Yooz apps consume the engine via the YoozEngineClient Swift Package — auto-discovery, auto-launch, REST + WebSocket clients in one.

Quick start

# Build
xcodegen generate
xcodebuild -project YoozEngine.xcodeproj -scheme YoozEngine \
  -configuration Debug -skipMacroValidation \
  -derivedDataPath build build

# Run
open build/Build/Products/Debug/Yooz\ Engine.app
curl http://localhost:19920/v1/health

-skipMacroValidation bypasses the per-machine trust prompt for MLXHuggingFaceMacros (required for headless / first-run CLI builds). -derivedDataPath build keeps the artifact next to the source tree instead of ~/Library/Developer/Xcode/DerivedData.

The full architecture, build variants, conventions, and module specs live in AGENTS.md.

The Infinite long-context module API, RAM tiers, model catalogue, and verification commands live in docs/INFINITE_MODULE.md.

Models

Yooz Engine expects model weights from HuggingFace. The Yooz-Labs published checkpoints live at huggingface.co/YoozLabs:

• YoozLabs/Qwen3-ASR-1.7B-Swift — multilingual ASR with Swift-friendly tokenizer
• YoozLabs/Yooz-Quality-v2-Qwen3.5-0.8B-LoRA — STT cleanup LLM (proofread + rewrite)
• More coming as we ship them.

All checkpoints are Apache 2.0. The engine code itself is source-available under PolyForm Shield 1.0.0 (see below).

License

The engine source code is licensed under PolyForm Shield 1.0.0. You can:

• Read, fork, modify, and use it for any purpose except building a competing product.
• Embed it in apps that aren't direct Yooz Engine substitutes.
• Contribute back via PRs.

You cannot offer a "managed Yooz Engine" or a re-skinned commercial fork.

Compiled .app binaries distributed via GitHub Releases are also under PolyForm Shield. Model weights on HuggingFace are Apache 2.0 (a separate, more permissive license — open weights are a deliberate choice).

For commercial-use or dual-license inquiries: [email protected].

Contributing

PRs welcome. Sign your commits with Signed-off-by: Your Name <[email protected]> (DCO style); see CONTRIBUTING.md for the full guide.

Status

Active development. Version 0.6.0 is the modular thin-client architecture (Phase 5). The engine is the heart of the Yooz ecosystem.

---

Part of the Yooz ecosystem. Sovereign Intelligence.