Building a 100% Local AI Therapy Notes App (Zero Cloud, Zero Training)
Nov 17, 2026Most AI therapy note tools pitch HIPAA compliance and "we don't train on PHI." Read the fine print, though, and many still reserve the right to train on de-identified or aggregated session data—or send sensitive transcripts through third-party cloud APIs. For therapists, school counselors, and clients, "trust our privacy policy" is a tough sell.
So we started building Minntelligence Notes: a 100% local-only desktop app where session notes never leave your machine.
How v1 Works
Built with Tauri v2, Rust, and llama.cpp (with Metal GPU acceleration on Apple Silicon), the app runs open-weight models directly on-device—meaning it works with Wi-Fi turned off.
- Paste anything, chat immediately: The app opens straight into a split Pasted Notes / Transcript + Chat workspace. Drop in rough bullet points, intake notes, or an entire 100k+ token transcript (automatically handled via local map-reduce chunking with a live progress bar) and ask questions about it.
- One-click clinical organization: Turn messy notes into structured SOAP, DAP, or BIRP notes, Summarize, Clean Up & Organize, or extract Themes & Follow-ups in one click.
- Starter model + Model Manager: Comes ready with a lightweight starter GGUF model for instant testing, plus an in-app catalog to download larger local models (like
Llama 3.2 3BorQwen 2.5 7B) when you want stronger clinical nuance. - Fast export: Copy any prompt or response, Export
.mddirectly to Downloads, or Export PDF cleanly for your EHR.
Why Local-First Matters
When inference stays on the clinician's laptop (or on dedicated, zero-retention hardware for a group practice or school district), you eliminate third-party AI subprocessors altogether—making client consent, HIPAA, and FERPA (NDPA) conversations dramatically simpler.
This is just v1—next up is local on-device audio transcription (Whisper) and deeper workflow templates.