Everything above is a claim. This is the receipt: the engines, the formats and the exact reason the app can run dark. No magic — a lot of carefully chosen, open-source, on-device code.
- Compute
- 100% on-device — Apple Metal + CoreML acceleration
- Network
- Nothing you open, type or export leaves the Mac; models download on demand and subscriptions send one small licence ping every 7 days
- Storage
- Models live in ~/Library/Application Support/LocalScribe
- Memory
- Model tasks pre-check ≈8 GB of free memory and refuse to start below it
- Silicon
- Apple Silicon (M1 or later) · macOS 14 Sonoma+
- Shell
- Tauri 2 + Rust core, React UI — no Electron, no bundled Chromium
- Retrieval
- ONNX embeddings on CoreML + a local vector index over your own document set
- Licensing
- ES256 (ECDSA P-256) JWT verified offline against a built-in public key; on Lifetime, activation is the only network call
● DenoiseDeepFilterNet3, 48 kHz float32
● Diarizepyannote voiceprints profiled on the cleaned audio, before the words
● Transcribewhisper.cpp · Metal · large-v3-turbo q8_0 · speaker labels merged in
● Clean & formatllama.cpp · GGUF Q4_K_M · glossary-constrained
● IndexCoreML embeddings → local vector store, built in the background
● Exportdocx / md / txt / pdf, from the raw or the cleaned transcript
Prefer to verify rather than read? Watch the network with Little Snitch or LuLu, or run the whole app inside a firewall-blocked profile. Both are the intended way to use it.