FerriScribe listens to your doctor-patient conversations and generates structured SOAP notes using a local AI model of your choice that runs entirely on your computer. No cloud APIs. No data transmission. No subscriptions. HIPAA-compliant by design — because nothing ever leaves your device.
System prompt systematically optimized and benchmarked across 21 real clinical cases. Achieves 90%+ schema compliance with zero fabricated patient data — every field defaults to "Not discussed" unless explicitly stated in the transcript.
All processing on-device. SQLCipher-encrypted database (AES-256). Audio recordings encrypted at rest with AES-256-GCM. No cloud APIs, no data transmission. HIPAA-compliant by architecture.
Records the consultation, transcribes speech to text, and generates a structured SOAP note — ready for your review before the next patient sits down.
Context templates for visit types, custom vocabulary find/replace rules, 6 letter audiences (Patient, Insurance, Tax Authority, Specialist, Employer, Legal) plus custom audiences. Generate referrals, clinical letters, and synopses with per-document custom prompts.
Audit every line of code. Fork it, contribute, or customize it for your clinic. No black boxes, no vendor lock-in, no future price hikes.
macOS, Linux, and Windows. Runs on any capable machine with 16GB+ RAM — from a MacBook Air to a clinic workstation.
No per-seat licenses. No API fees. No cloud subscriptions. Download it, run it, own it. One less line item on your practice budget.
Pair laptops to a central office server over LAN or Tailscale. Bidirectional content sync keeps SOAP notes, transcripts, and recordings consistent across all devices — with per-field last-write-wins merge and automatic conflict resolution.
Proven on Apple Silicon Mac Studio with 192GB unified memory. Runs quantized models (Qwen 3.6 35B A3B) via LM Studio — no GPU server, no special equipment needed.
7,122 bundled BC MSP ICD-9 diagnostic codes with automatic SOAP validation. Off-list codes flagged as amber chips for clinician review. Inverted-index lookup for instant code selection.
Ingest clinical documents into a local vector store. Retrieval-Augmented Generation combines BM25, vector similarity, and graph search. Multi-step agentic workflows with tool use for complex clinical queries.
Export SOAP notes as PDF, DOCX, or FHIR R4 — the healthcare interoperability standard. Share structured clinical documents with any EHR system. Audio recordings exportable as standard WAV files.
Space to record/stop, Cmd/Ctrl+Enter to generate SOAP notes. RSVP speed reader for rapid document review at configurable WPM with per-section filters. Drag-and-drop condition chip reordering.
Encrypted database with OS keychain integration. PHI-redacted log export for safe technical support. Soft-delete recordings with undo toast. Crash recovery for interrupted encryption operations.
Persistent condition tracking with drag-and-drop reorder. Real-time sync across paired devices via SSE. Tombstone-based conflict resolution ensures consistency across your practice.
Hit record before the consultation. FerriScribe captures audio and transcribes it locally with high-quality speech-to-text via Whisper.
The local AI model analyzes the transcript on-device — extracting clinical findings, ICD codes, differential diagnoses, assessments, and plans.
A structured SOAP note appears in seconds. Review, edit, and copy it into your EMR. Done before your next patient.
FerriScribe runs entirely on your local machine — no cloud APIs, no data transmission, no third-party servers. Patient conversations never leave your device. This isn't a privacy policy you have to trust; it's an architecture you can verify. Every line of code is open source and auditable.
FerriScribe isn't a prototype — it's running daily in a real family practice, generating SOAP notes from actual patient encounters. Here's the reference deployment:
Native Apple Silicon support with MLX optimization. Runs great on MacBook Air, MacBook Pro, and Mac Studio with M-series chips.
Full NVIDIA CUDA and AMD ROCm support. Run it on your workstation, a home server, or a dedicated clinic machine.
DirectML and CUDA support. Works on any capable Windows laptop or desktop with sufficient RAM for local inference.
The only requirement: a computer with enough RAM to run local models (16GB+ recommended).