Yes, there is a therapy-note AI that doesn't use the cloud — here's how it works

Why most clinicians assume this doesn't exist

Until recently, they were right. Speech-to-text good enough for emotional, overlapping, soft-spoken therapy dialogue required a large cloud model. Note drafting good enough for a clinician to edit rather than rewrite required an even larger one. Running both on a laptop was absurd — consumer GPUs had 8 GB of VRAM, and a usable model at full precision needed 50 GB or more.

Three things changed between 2023 and 2026:

1. Apple unified memory. M-series Macs share RAM between CPU and GPU, so a 14 B-parameter model occupies system memory directly rather than requiring a dedicated GPU card. A 24 GB M3 Pro has more addressable GPU memory than a $1,500 Windows laptop.
2. 4-bit quantization. Techniques like AWQ and GPTQ compress a 14 B-parameter model from ~28 GB down to ~8 GB with ≤2% quality loss on instruction-following benchmarks. The Qwen 2.5 family specifically publishes official 4-bit weights. The model goes from "cloud-only" to "fits in RAM."
3. whisper.cpp + Core ML acceleration. whisper.cpp runs OpenAI's Whisper models on CPU, and since 2024 with a Core ML path that lights up the Apple Neural Engine. Real-time factor on M2 is under 1×, meaning a 50-minute session transcribes in less than 50 minutes — in practice, 3–4 minutes.

None of these are proprietary to us. They are open-source components any engineer can audit or reproduce. What we did was stitch them together with a clinically-aware prompt template, a tamper-evident log, and a macOS distribution that enforces the no-socket guarantee.

What "without cloud" actually looks like on your machine

When you drop a session audio file into TherapyDraft:

• The file stays on disk in ~/Library/Application Support/TherapyDraft/audio/. It is not copied to a vendor server. It is not uploaded anywhere. It is readable only by your user account.
• The transcription pass reads that file, runs whisper.cpp on the Apple Silicon GPU, and writes the transcript to an adjacent file. No HTTP requests. No DNS lookups.
• The drafting pass reads the transcript, runs Qwen 14B via MLX locally, and writes the draft to a third adjacent file. Again, no network activity.
• The tamper-evident JSON-Lines inference log appends a row for each pass (model ID, prompt hash, output hash, timestamp, device ID). The log is local; no one but you can read it.

You can run sudo lsof -i -n -p $(pgrep TherapyDraft) during any of this and see zero open sockets besides a Stripe license check at startup. We're not arguing about it — you can check.

Is the output really clinical-quality without cloud-scale parameters?

For SOAP/DAP/BIRP/GIRP drafting from a transcript, a 14 B-parameter model with a well-crafted prompt and five of the clinician's own example notes as few-shot anchors produces output that is, in our internal benchmarking, a ~1-minute edit away from sendable. A 70 B cloud model will sometimes produce a draft that needs only 30 seconds of edit instead of 60 — but neither number is "zero," and for many clinicians the 30 seconds of time saved does not justify shipping the audio to a vendor.

The important honesty: AI scribe output of any architecture is a first draft. Reading and correcting it is a clinical-responsibility step, not a workflow inconvenience. See the pricing-comparison page for the side-by-side on cost-per-session and quality trade-offs.

What you still need to bring

A recent M-series Mac (M1 or later) with 16 GB of RAM as the recommended floor; 24 GB if you want sub-3-minute turnarounds. macOS 14 Sonoma or later. About 10 GB of free disk for the bundled models. No special audio hardware — the Mac's built-in microphone handles in-person sessions fine; a USB lavalier helps for the first few weeks while you calibrate volume.

Related questions

Further reading

• On-device therapy note generator: the latency table
• Private AI therapy scribe: the sandbox-entitlement layer
• HIPAA AI SOAP note: the no-BAA argument
• Join the private beta