Installation

macOS (Current Release)

Download the .dmg file from the download page. Open the disk image and drag ClinicalFlow into your Applications folder. On first launch, macOS may ask you to confirm you want to open an app from an identified developer — click Open.

Supported on macOS 11.0 (Big Sur) or later. Optimized for Apple Silicon (M1, M2, M3, M4).

Windows & Linux

Windows and Linux builds are in development. Currently, ClinicalFlow is macOS-only. Sign up for notifications on the download page to be notified when these platforms become available.

First Launch & PIN Setup

When you open ClinicalFlow for the first time, you’ll be asked to create a numeric PIN (4–8 digits). This PIN encrypts all clinical data stored on your device using AES-256 encryption.

You’ll enter this PIN every time you open the app. ClinicalFlow never stores your PIN in plaintext — only an Argon2id cryptographic hash is saved for verification.

After creating your PIN, the Welcome Wizard guides you through initial setup.

Welcome Wizard

The Welcome Wizard runs once after PIN creation and walks you through a 5-screen setup flow:

After completing the wizard, it never appears again. The welcomeCompleted flag is saved in your configuration. To re-run the setup, delete your auth.json file (see Log Files for file locations).

Demo Mode

ClinicalFlow includes two built-in demo encounters to help you explore the app without a real patient visit.

Accessing Demo Mode

Triple-click the Help button (3 clicks within 400ms). A picker modal appears with two demo scenarios.

Medical Demo

A multi-problem internal medicine encounter with James Robinson (patient), Dr. Patel (physician), and Maria (medical assistant). Covers Type 2 diabetes, hypertension, peripheral neuropathy, and left knee osteoarthritis. The transcript plays back word-by-word with waveform animation (~24 entries).

Dental Demo

A comprehensive dental encounter with Ms. Ramirez (patient) and Dr. Chen (dentist). Covers MOD caries, endodontic failure, impacted wisdom tooth, and a 4-phase treatment plan. The transcript loads instantly (15 entries), the dental chart auto-populates with findings, and the dental_general template is auto-selected.

Demo mode is a great way to test different note templates, dental charting, medical coding, and the verification system before using ClinicalFlow in a real encounter.

System Requirements

Online Mode (Minimum)

Offline Mode (Recommended)

Recording Basics

The record button is the large circular button at the bottom center of the app. Press it or hit Space to start recording. The button pulses red and a timer counts up in the header.

Controls

During Recording

• The timer in the header counts up (MM:SS format)
• The waveform visualizer animates at 60 FPS to show audio input level
• Transcript entries appear in real time in the left panel
• Each entry shows the speaker name, timestamp, and transcribed text
• The transcript auto-scrolls to the latest entry (can be toggled off in Settings)

Maximum Session Duration

Sessions are limited to 4 hours of continuous recording. When this limit is reached, recording stops automatically.

Crash Recovery

ClinicalFlow writes audio data to disk continuously during recording (flushing every ~10 seconds). If the app crashes mid-recording, your audio is preserved up to the last flush. On restart, ClinicalFlow recovers the session automatically.

Online Transcription (Deepgram)

Online mode uses Deepgram Nova-3 Medical, a cloud-based speech recognition model optimized for clinical encounters. It provides the highest accuracy and real-time speaker diarization.

Setup

1. Sign up at deepgram.com (free tier includes $200 of credit)
2. Create an API key in your Deepgram dashboard
3. Open ClinicalFlow Settings → Transcription → paste your API key → click Save

The connection indicator in the header will show Online — connected when the key is valid.

How It Works

Audio is captured from your microphone, encoded as 16-bit PCM at 16 kHz mono, and streamed to Deepgram over an encrypted WebSocket connection (WSS). Deepgram returns transcription results with word-level timestamps and speaker labels. A keepalive signal is sent every 8 seconds, and utterance boundaries are detected after 1,500ms of silence.

Keyterm Boosting

ClinicalFlow sends up to 100 medical terms to Deepgram as keyterm parameters, drawn from the corrections dictionary and medical dictionary. This improves recognition accuracy for clinical terminology.

Speaker Diarization

Deepgram automatically identifies different speakers in the audio stream. ClinicalFlow maps these speaker IDs to the speakers you’ve defined in the sidebar.

Offline Transcription (Whisper)

Offline mode uses a local Whisper Small model (465 MB, multilingual) running via a whisper-server HTTP sidecar. No audio ever leaves your machine.

How It Works

Audio is captured in 3-second chunks with 0.5-second overlap. Each chunk is sent as a 16 kHz mono WAV to the local whisper-server via HTTP POST. The server keeps the model loaded in memory between requests for fast inference. Chunks below a 200 RMS silence threshold are skipped to avoid processing silence.

Model Details

Microphone Setup

ClinicalFlow uses your system’s default audio input device. Audio is captured at 16 kHz, 16-bit signed mono via the native audio stack (cpal on Rust).

For best results, use an external USB microphone positioned between you and the patient. Built-in laptop microphones work but capture more ambient noise.

On macOS, you’ll need to grant microphone permission on first launch: System Settings → Privacy & Security → Microphone → enable ClinicalFlow.

Languages (37)

ClinicalFlow supports 37 languages across the full pipeline: transcription, medical term corrections, note generation, and date formatting.

Correction Dictionaries

Medical term corrections are available for 6 languages, totaling 204 patterns:

Languages without dedicated correction dictionaries use the base transcription engine output without post-processing.

Medical Term Corrections

Speech-to-text engines frequently garble medical terminology. ClinicalFlow applies a post-processing step after transcription that fixes common errors using 204 regex patterns across 6 languages.

What Gets Corrected

Custom Dictionary

Corrections are loaded from language-specific JSON files in the app’s data directory (corrections.json for English, corrections-es.json for Spanish, etc.). You can edit these files to add corrections specific to your specialty:

[
  { "pattern": "\\bmetforming\\b", "flags": "gi", "replacement": "metformin" },
  { "pattern": "\\byour-custom-term\\b", "flags": "gi", "replacement": "CorrectTerm" }
]

ClinicalFlow reloads the corrections file on each app launch. Changes take effect on the next recording.

Hallucination Filtering

Whisper occasionally produces false positive text output from silence or background noise. ClinicalFlow filters these automatically before they reach the transcript.

What Gets Filtered

• Whisper meta-tags: [blank_audio], (silence), (blank audio), (no audio)
• Background sounds: (music), (laughter), (coughing), (breathing), (footsteps), and similar
• YouTube/social media artifacts: “subscribe”, “like and subscribe”, “thank you for watching” (training data leakage from Whisper)
• Generic filler: “thank you.”, “goodbye.”, “the end.”, single-word outputs
• Structural patterns: Fully parenthesized strings, repeated words/phrases, empty output

Medical Vocabulary Conditioning

ClinicalFlow conditions the Whisper model with a prompt containing 55+ medical terms to improve recognition of clinical terminology. This prompt is passed to whisper-server via the --prompt flag.

Term Categories

Speaker Management

The sidebar’s Speakers section lets you define who is in the encounter. Each speaker gets a name, a role, and a unique color that appears in the transcript.

Adding a Speaker

Click the + button in the Speakers section. Enter a name (e.g., “Dr. Martinez”) and select a role. ClinicalFlow automatically assigns a color.

Switching the Active Speaker

Click a speaker’s name in the sidebar to make them the active speaker. All subsequent transcript entries will be attributed to this speaker until you switch again or auto-detection switches for you.

Removing a Speaker

Click the × button next to a speaker’s name. Existing transcript entries from that speaker are preserved.

Roles & Colors

Each speaker can be assigned one of three roles. The role determines how the speaker’s words are categorized in the generated clinical note:

Each speaker has a unique color in the transcript, displayed as a left border on each entry for easy visual scanning.

Auto-Detection

When Auto-detect speakers is enabled (Settings → Recording), ClinicalFlow uses two methods to automatically switch speakers:

• Diarization data (online mode) — Deepgram provides speaker IDs with each transcript segment, and ClinicalFlow maps these to your defined speakers.
• Silence detection (both modes) — After 1.5 seconds of silence, ClinicalFlow may switch the active speaker. This works best in clear turn-taking conversations.

Auto-detection is not perfect. You can always manually click a speaker in the sidebar to correct it.

Cloud AI (Claude Haiku 4.5)

Cloud AI uses Anthropic’s Claude Haiku 4.5 (claude-haiku-4-5-20251001) to generate clinical notes. It provides the highest quality output with the most natural clinical language.

Setup

1. Sign up at console.anthropic.com
2. Create an API key (starts with sk-ant-)
3. Open ClinicalFlow Settings → AI Note Generation → select Cloud AI
4. Paste your API key and click Save
5. Click Test to verify the connection

Technical Details

Cost

A full online encounter (Deepgram transcription + Claude note generation with verification) costs approximately $0.10–$0.15 per encounter for a 15-minute visit. This is separate from your ClinicalFlow subscription — you pay Deepgram and Anthropic directly through your own API keys.

How It Works

When you click Generate Note (or press G), ClinicalFlow sends the full transcript to Claude along with a detailed clinical prompt that specifies the note template, evidence-extraction rules, safety constraints, and output structure. Claude’s response streams back in real time — you see the note being written section by section.

Ollama (Local AI)

Ollama runs AI models entirely on your local machine. No data leaves your computer. It’s free, open-source, and supports dozens of models.

Installation

1. Download Ollama from ollama.com/download
2. Install and launch it (it runs in the background)
3. Open Terminal and pull a model: ollama pull llama3.1:8b
4. In ClinicalFlow Settings → AI Note Generation → select Ollama
5. Click Test to verify the connection, then Refresh to load available models.

Recommended Models

Configuration

Default server URL: http://localhost:11434. ClinicalFlow includes 3 automatic retries for transient failures.

Rule-Based Fallback

If both Cloud AI and Ollama are unavailable, ClinicalFlow falls back to a rule-based note generator. This deterministic system extracts information using pattern matching: vital sign patterns, medication mentions, symptom descriptions, and exam findings.

Rule-based notes are less polished than AI-generated ones but provide a reliable baseline that captures key data points. Select it explicitly in Settings → AI Note Generation → Rule-Based.

Two-Pass Verification

ClinicalFlow’s verification system runs a second AI pass after generating a note. This auditor checks every claim against the original transcript.

The verification pass runs at temperature 0.1 for maximum precision. Toggle in Settings → AI Note Generation → Verification pass.

Clinical Prompt System

ClinicalFlow uses a carefully engineered prompt to guide AI note generation. The prompt includes:

• Template specification — Exact section headers and structure for the selected template (from all 20 built-in templates)
• Evidence-extraction rules — Every line must trace back to something explicitly stated in the transcript
• Omission logic — If a system or exam component was not discussed, it is omitted (not listed as “not assessed”)
• Pertinent negatives — Only negatives the patient explicitly denied are included
• Medication categorization — Separates new prescriptions, continued medications, and dosage changes
• Multilingual output — For non-English languages, an OUTPUT LANGUAGE directive generates clinical content in the target language while keeping section headers in English for the note parser
• Dental chart integration — When a dental template is selected, dental chart findings are serialized into the prompt
• Medical/dental coding — When auto-coding is enabled, a separate coding prompt generates ICD-10, CPT, E&M, or CDT codes from the note content

Template Overview

ClinicalFlow includes 20 built-in note templates across 5 categories, plus the ability to create custom templates. Select a template from the dropdown (#formatSelector) in the sidebar, or change your default in Settings.

General Templates (3)

Behavioral Health Templates (2)

Specialty Templates (10)

Dental Templates (5)

Dental templates automatically enable the interactive dental chart in the sidebar.

Custom Templates

Click + Custom Template in the template dropdown to create your own. Provide a name, define section names, and optionally write a custom AI prompt. Custom templates are stored in the ms-custom-templates config key as a JSON array.

Interactive Dental Chart

When any dental template is selected, a Dental Chart section appears in the sidebar with an “Edit Chart” button. Clicking opens a full-screen SVG dental chart modal.

The chart supports both 32-tooth adult (permanent) and 20-tooth primary (deciduous) dentition with a toggle switch. Teeth use ADA Universal numbering: adult teeth 1–32, primary teeth A–T.

Tooth States (8)

Click any tooth to open a two-tier popup. Select from 8 color-coded clinical states:

Surface Marking

For states that affect specific surfaces (decay, restored, fracture), a second tier of surface buttons appears after selecting the state:

Anatomical validation prevents impossible surface combinations. Click outside the popup to save your selection.

AI Integration

Dental chart findings are automatically serialized into the AI prompt during note generation. The format includes dentition type, tooth number, state, and affected surfaces.

Response Parsing

When the AI generates a note mentioning dental findings, ClinicalFlow can parse findings back to the chart. Acronym protection prevents false surface detection for 18 common dental abbreviations: FPD, RPD, ZOE, FGC, PFM, MTA, SDF, BOP, CAL, TMD, TMJ, GBI, OHI, DI, CI, CEJ, PDL, and IPR.

Export

The dental chart SVG and findings table are included in PDF and text exports when enabled in Settings (ms-settings-dentalChartInExport and ms-settings-dentalFindingsInExport).

Periodontal Charting Overview

ClinicalFlow includes a full periodontal charting mode that overlays on the dental chart. When a dental template is active, a “Perio Mode” toggle appears in the dental chart modal. Enabling perio mode transforms the chart into a 6-point probing depth entry system.

Accessing Perio Mode

• Select any dental template (General Dental, Periodontal Exam, etc.)
• Open the dental chart modal via Edit Chart
• Toggle the Perio Mode switch at the top of the chart

What Perio Mode Adds

• 6-point probing depths per tooth (MB, B, DB, ML, L, DL)
• Bleeding on probing (BOP) markers per site
• Recession measurements per site
• Mobility grades (I, II, III) per tooth
• Furcation involvement grades (I, II, III) for multi-rooted teeth
• Clinical attachment level (CAL) auto-calculated from probing depth + recession

All perio data is serialized into the AI prompt during note generation and included in PDF/text exports.

Voice-Driven Entry

Probing depths and perio findings can be entered by voice during a live recording. The perio voice parser recognizes natural dictation patterns and maps them to specific teeth and sites.

Supported Voice Commands

Voice-entered data populates the perio chart in real time. You can also click individual sites on the chart to enter or correct values manually.

Measurements & Data Points

Each tooth in perio mode tracks the following data points:

Color Coding

Probing depths are color-coded for quick visual assessment:

Documentation Completeness Scoring

ClinicalFlow automatically calculates a documentation completeness score for each perio charting session. This helps ensure thorough documentation before generating notes or exporting.

How Scoring Works

The score is a percentage (0–100%) based on the ratio of documented data points to expected data points for the teeth present in the chart. Missing teeth are excluded from the calculation.

Score Display

The completeness score appears as a badge in the perio chart header. Scores below 80% display a warning suggesting the provider complete additional measurements before generating the note.

Automatic Billing Codes

ClinicalFlow automatically generates billing codes from the generated note content when auto-coding is enabled (ms-settings-autoCoding). Codes appear in a read-only panel (#codingPanel) below the note sections with colored confidence badges.

Confidence Levels

Codes are read-only with a copy-to-clipboard option. No direct editing of suggested codes in the UI.

Medical Coding (ICD-10 + CPT)

For medical (non-dental) templates, ClinicalFlow generates:

• Up to 8 ICD-10-CM diagnosis codes with confidence levels
• Up to 4 CPT procedure codes with confidence levels
• E&M level assessment (Level 1–5) based on number of problems addressed, data complexity, and risk of morbidity/mortality

The system uses only real, valid ICD-10-CM and CPT codes.

Dental Coding (CDT + Dental ICD-10)

For dental templates, ClinicalFlow generates CDT D-codes and dental-specific ICD-10 codes. The dental coding engine enforces 4 strict guardrails:

1. Strict Domain Gate

Only CDT D-codes and dental ICD-10 codes are allowed. CPT codes (00100–99499), E&M levels, and non-dental codes are strictly forbidden. Every procedure code must begin with “D”.

2. Depth & Severity Escalator

Never defaults to the shallowest/generic code. If “approaching pulp” or “deep caries” is documented, codes escalate to K04.0/K04.1 rather than K02.51. If “bone loss” or “deep pockets >5mm” is documented, periodontal codes escalate from K05.10 to K05.311+.

3. Structural vs. Positional Trauma

Every injury is classified into exactly one track:

• Track A (structural failure): cracked/fractured → K03.81, S02.5XXA, D2740, D7140
• Track B (positional displacement): loosened/luxated → S03.2XXA, M26.30, D7270, D4921

Codes from both tracks are never assigned to the same tooth.

4. Etiology of Absence

For every missing tooth, the reason is determined. Default is acquired (K08.111–K08.409) unless the transcript explicitly states “congenitally absent”, “agenesis”, or “hypodontia” → then congenital (K00.0).

Audit Flags

After billing codes are generated, ClinicalFlow runs an audit pass that checks for common documentation and coding issues. Audit flags appear as collapsible alerts below the coding panel.

What the Audit Checks

Interacting with Flags

Each flag is collapsible — click to expand for details and the specific code(s) involved. Flags are informational only and do not modify the suggested codes. They help providers review and correct coding before submission.

Clinical Warnings

Separate from audit flags, ClinicalFlow generates clinical warnings when the note content suggests potential safety or quality concerns. Warnings appear as highlighted alerts in the coding panel.

Warning Categories

Inline Editing

After note generation, each section is editable directly in the note panel. Click on any section content to make changes. Edits are saved to the active session automatically.

PDF Export

Press E or click the Export PDF button. ClinicalFlow uses a native Swift html2pdf sidecar (WebKit createPDF) to generate a professional PDF with header, formatted sections, disclaimer footer, and optional dental chart.

PDFs are saved to the app’s exports directory with the filename ClinicalFlow_Note_YYYY-MM-DD_HH-MM.pdf.

Text Export

Press T to download the transcript as a plain text file. A native save dialog lets you choose the location. The default filename is ClinicalFlow_Transcript_YYYY-MM-DD.txt.

Copy to Clipboard

Click the Copy button in the note panel header to copy the entire note as plain text. Paste it directly into your EHR (Epic, Cerner, Athena, or any system that accepts text input).

Export Actions

ClinicalFlow provides configurable export actions that appear in the note panel toolbar after note generation. You can enable or disable individual actions to streamline your workflow.

Available Actions

Action visibility is managed in Settings. Disabled actions are hidden from the toolbar entirely, keeping the interface clean.

Session Archiving

When you click New Session (or press N), ClinicalFlow archives the current encounter. The archive includes the full transcript, generated note, audio recording, dental chart data, coding results, and all session metadata.

Archives are named YYYY-MM-DD_HH-MM_PatientName, encrypted with AES-256-GCM, and sorted newest-first in the archive browser. Audio is stored as a separate unencrypted WAV file.

EHR Integration

ClinicalFlow can export clinical notes to Electronic Health Record systems using the HL7 v2 messaging standard. This allows direct transmission of generated notes into your EHR without manual copy-paste.

Setup

Configure EHR integration in Settings → Advanced → EHR Integration:

How It Works

• After note generation, click Send to EHR in the export toolbar
• ClinicalFlow packages the note into an HL7 v2 ORU^R01 message
• The message includes patient identifiers, note sections, coding data, and dental findings
• The HL7 message is sent to the configured endpoint

PMS Bridge

The Practice Management System (PMS) Bridge connects ClinicalFlow to your dental or medical practice management software for bidirectional data exchange.

Supported Systems

Configuration

Configure PMS Bridge in Settings → Advanced → PMS Bridge:

Use the Test Connection button to verify connectivity before saving. A successful test confirms ClinicalFlow can reach the PMS database.

Insurance Narrative Generation

ClinicalFlow can generate insurance-ready narratives from the clinical note and billing codes. These narratives are formatted for insurance claim submissions and pre-authorization requests.

What the Narrative Includes

• Clinical justification — summarizes the patient’s condition and why the treatment is medically/dentally necessary
• Supporting documentation — references specific findings from the examination
• Procedure rationale — links each procedure code to the clinical indication
• Diagnosis correlation — maps ICD-10 or dental ICD-10 codes to the documented findings

Generating a Narrative

After note generation with coding enabled, click Generate Narrative in the export toolbar. The AI processes the note content, billing codes, and dental chart data to produce a cohesive narrative. The narrative can be copied to clipboard or exported as part of the PDF.

General & Advanced Tabs

The Settings drawer is organized into two tabs accessible via a segmented control at the top:

Click a tab to switch views. Your selected tab is remembered for the duration of the session.

Appearance

Transcription Settings

AI Engine Settings

Recording Settings

EHR Integration Settings

Found in Settings → Advanced tab → EHR Integration card.

PMS Bridge Settings

Found in Settings → Advanced tab → PMS Bridge card.

Language Settings

Found in Settings → Advanced tab → Language card.

Security Settings

Keyboard Shortcuts

Encryption at Rest

ClinicalFlow encrypts all files containing protected health information using AES-256-GCM with keys derived from your PIN via PBKDF2-HMAC-SHA256 (100,000 iterations).

What Is Encrypted

Each encrypted file has a unique random salt (16 bytes) and nonce (12 bytes) prepended to the ciphertext. Total overhead: 28 bytes per file. Writes are atomic (write to .tmp, then rename) to prevent corruption.

PIN & Lock Screen

The PIN serves two purposes: authentication (proving you’re authorized) and key derivation (your PIN is the cryptographic key that unlocks encrypted files).

• PIN is verified using Argon2id hashing (memory-hard, resistant to brute force)
• Only the hash is stored (in auth.json), never the PIN itself
• PIN length: 4–8 numeric digits
• No lockout after failed attempts — the encryption itself is the protection
• Auto-lock default: 5 minutes of inactivity (60 seconds when page is hidden)
• Auto-lock is deferred during active recording to prevent data loss

Network Requests

HIPAA Considerations

ClinicalFlow implements HIPAA technical safeguards: encryption at rest (AES-256-GCM), unique user authentication (PIN), automatic logoff (auto-lock), audit logging (sanitized, no PHI), transmission security (TLS for all network requests), Content Security Policy headers, and localhost-only Whisper binding.

For full HIPAA compliance, your organization also needs Business Associate Agreements (BAAs) with Deepgram and Anthropic if using online features, plus organizational policies covering training, incident response, and access management.

Data Handling Summary

Plans & Billing

ClinicalFlow uses Supabase for authentication and Stripe for payment processing.

Subscription States

Manage your subscription through the in-app billing portal, which opens Stripe’s management interface.

Microphone Issues

“No microphone found”

Ensure a microphone is connected. Check System Settings → Sound to confirm it’s recognized. Try unplugging and replugging USB microphones.

“Microphone access denied”

macOS: System Settings → Privacy & Security → Microphone → enable ClinicalFlow.

Transcription Issues

Poor transcription accuracy

• Use an external USB microphone instead of built-in laptop mic
• Reduce ambient noise (fans, HVAC, other conversations)
• Speak clearly and at a normal pace
• Check that the correct language is selected in Settings

Deepgram connection failed

• Verify your API key in Settings
• Check your internet connection
• Check status.deepgram.com for outages
• Switch to offline mode as a workaround

Ollama Issues

“Ollama not connected”

• Make sure Ollama is running: ollama list in Terminal
• If not installed, download from ollama.com
• Check the server URL in Settings (default: http://localhost:11434)

Out of memory

Switch to a smaller model (e.g., llama3.1:8b instead of llama3.1:70b) or use Cloud AI.

Note Generation Issues

Note seems incomplete or inaccurate

• Ensure the transcript is complete before generating
• Enable the verification pass to catch omissions
• Check that speakers are correctly labeled
• Try a different template that better matches the encounter type

Cancel a stuck generation

The Generate button becomes a Cancel button during generation. Click it to abort. ClinicalFlow uses an AbortController to cleanly cancel API calls. Timeout is 120 seconds.

Export Issues

PDF export shows blank page

Try exporting again. If the issue persists, use Copy to Clipboard as an alternative.

Save dialog doesn’t appear

Check that ClinicalFlow has file access permissions in System Settings → Privacy & Security → Files and Folders.

Crashes & Recovery

App crashed during recording

Your audio is safe. ClinicalFlow writes to disk every ~10 seconds. On restart, the session is recovered automatically.

App crashed during note generation

The transcript is preserved. Regenerate the note from the saved transcript after restart.

Low memory warning

ClinicalFlow monitors available system memory. Warnings appear at <2 GB free. Close other applications or switch to a smaller model.

Log Files

ClinicalFlow writes structured logs for debugging. Logs never contain patient data.

File Locations

Log files are named clinicalflow-YYYY-MM-DD.log, rotated daily, and auto-cleaned after 30 days.

System Architecture

ClinicalFlow is a desktop application built with Tauri v2. The frontend is vanilla HTML/CSS/JavaScript (17 modules, ~6,065 lines) running in a native webview. The backend is Rust (6 modules, ~2,200 lines), handling audio capture, file encryption, and system integration. No JavaScript frameworks are used.

Component Overview

┌─────────────────────────────────────────────────────┐
│                   Tauri Webview                      │
│                                                     │
│  ┌──────────────┐  ┌──────────────┐  ┌───────────┐ │
│  │ Transcript   │  │ Note Panel   │  │ Sidebar   │ │
│  │ Panel        │  │ (editable)   │  │           │ │
│  │              │  │              │  │ Speakers  │ │
│  │ Real-time    │  │ 20 templates │  │ Templates │ │
│  │ entries      │  │ Dental chart │  │ Settings  │ │
│  │ 1,300+ terms │  │ Coding panel │  │ Archive   │ │
│  └──────────────┘  └──────────────┘  └───────────┘ │
│                                                     │
│  17 JS modules — UI, prompts, templates, dental,   │
│  notes, coding, subscription, audio, settings       │
└─────────────────────┬───────────────────────────────┘
                      │  invoke() / listen()
                      │  (26 Tauri IPC commands)
┌─────────────────────┴───────────────────────────────┐
│                  Rust Backend                        │
│                                                     │
│  audio.rs     — cpal capture, whisper-server,       │
│                  hallucination filter, vocab prompt  │
│  storage.rs   — Encrypted sessions, config,         │
│                  archives, PDF export, logging       │
│  crypto.rs    — AES-256-GCM, PBKDF2 key derivation │
│  auth.rs      — Argon2id PIN, auto-lock, AppState   │
│  logging.rs   — Tracing, daily rotation             │
│  lib.rs       — App setup, command registration     │
└─────────────────────┬───────────────────────────────┘
                      │
          ┌───────────┼───────────┐
          ▼           ▼           ▼
┌──────────────┐ ┌─────────┐ ┌──────────────────┐
│whisper-server│ │html2pdf │ │  Local Storage    │
│ (3.44 GB)    │ │ (62 KB) │ │                  │
│              │ │ Swift   │ │ auth.json        │
│ ggml-small   │ │ WebKit  │ │ config.json  🔒  │
│ 465 MB model │ │ PDF gen │ │ active.json  🔒  │
│ HTTP sidecar │ │         │ │ archive/*.json🔒 │
│ 127.0.0.1    │ │         │ │ archive/*.wav    │
└──────────────┘ └─────────┘ │ logs/*.log       │
                             └──────────────────┘

External Services (online mode only):
┌───────────────┐ ┌───────────────┐ ┌──────────────┐
│ Deepgram API  │ │ Anthropic API │ │ Supabase     │
│ WSS streaming │ │ HTTPS (Claude)│ │ Auth + Stripe│
└───────────────┘ └───────────────┘ └──────────────┘

┌───────────────┐
│ Ollama        │
│ localhost     │
│ :11434        │
└───────────────┘

🔒 = AES-256-GCM encrypted with PIN-derived key

Data Formats

File Storage Structure

~/Library/Application Support/com.clinicalflow.ai/
├── auth.json                      PIN hash + welcome flag
├── config.json                    [ENCRYPTED] Settings & API keys
├── sessions/
│   ├── active.json               [ENCRYPTED] Current session
│   └── archive/
│       ├── 2026-02-23_18-45_PatientName.json  [ENCRYPTED]
│       └── 2026-02-23_18-45_PatientName.wav   Audio recording
├── exports/
│   ├── ClinicalFlow_Note_2026-02-23.pdf
│   └── ClinicalFlow_Transcript_2026-02-23.txt
└── logs/
    └── clinicalflow-2026-02-23.log

All session data is stored as JSON files (no database). Sessions include entries, speakers, note sections, dental chart state, and coding results.

How ClinicalFlow Compares

An honest comparison of where ClinicalFlow fits relative to other AI clinical documentation tools.

ClinicalFlow’s Strengths

• Offline mode. Most AI scribes are cloud-only. ClinicalFlow runs with zero internet.
• 20 specialty templates + dental & perio charting. Covers general, behavioral health, 10 specialties, 5 dental templates, full periodontal charting, and custom templates.
• Automatic billing codes with audit flags. ICD-10, CPT, E&M, and CDT coding with audit flags and clinical warnings.
• EHR & PMS integrations. HL7 v2 export to EHR systems, PMS Bridge for Dentrix/Eaglesoft/Open Dental, and insurance narrative generation.
• Transparent pricing. $25/month with 14-day free trial. Cloud API costs separate and pay-as-you-go.
• Data stays local. In offline mode, nothing leaves your machine. No ClinicalFlow servers.
• Open architecture. Bring your own API keys, choose your AI models.
• Encryption at rest. PIN-based AES-256 encryption is not standard in most scribe tools.

Where Others May Have an Advantage

• Deep EHR integration. Products like Nuance DAX, Abridge, and Suki offer embedded EHR integration. ClinicalFlow supports HL7 v2 export but does not embed directly inside EHR interfaces.
• Ambient listening. Some products run continuously without pressing a button. ClinicalFlow requires explicit start/stop.
• Multi-user and team features. Enterprise products offer multi-provider support, shared templates, and admin dashboards. ClinicalFlow is single-user.
• Mobile apps. Some competitors have iOS/Android apps. ClinicalFlow is macOS desktop-only.
• Cross-platform. ClinicalFlow currently runs only on macOS (Apple Silicon). Windows and Linux are planned.

Summary Table

Version History

v1.0.0 — Initial Release

February 2026

• Dual-mode transcription: Deepgram Nova-3 Medical (online) and Whisper Small (offline, 465 MB)
• AI note generation: Claude Haiku 4.5 (cloud), Ollama (local), and rule-based fallback
• 20 built-in note templates across 5 categories (General, Behavioral Health, Specialty, Dental, Preventive) plus custom templates
• Two-pass verification system for hallucination, contradiction, and omission detection
• Interactive dental chart: 32-tooth adult + 20-tooth primary, 8 states, 5-surface marking
• Periodontal charting: 6-point probing depths, BOP, recession, mobility, furcation, voice-driven entry
• Documentation completeness scoring for perio charting sessions
• Medical coding: ICD-10 (up to 8), CPT (up to 4), E&M level assessment
• Dental coding: CDT D-codes + dental ICD-10 with 4 guardrails
• Audit flags and clinical warnings for coding review
• Multi-speaker management with 3 roles (Physician, Patient, Other) and 1.5s silence detection
• Medical term post-processor with 204 correction patterns across 6 languages
• Whisper hallucination filtering and medical vocabulary conditioning (55+ terms)
• PIN-based authentication with AES-256-GCM encryption at rest (PBKDF2 100k iterations, Argon2id)
• Auto-lock after configurable inactivity (default: 5 minutes)
• EHR integration via HL7 v2 export
• PMS Bridge for Dentrix, Eaglesoft, and Open Dental
• Insurance narrative generation from notes and billing codes
• PDF export (Swift WebKit sidecar), text export, clipboard copy, configurable export actions
• Session archiving with transcript, note, audio, dental chart, and coding data
• Continuous WAV save (~10s intervals) for crash recovery
• Structured logging with PHI sanitization (30-day retention)
• Settings drawer with General/Advanced tabs and squircle card layout
• Dark and light themes
• 37-language support (full pipeline: transcription + corrections + notes + UI)
• 1,300+ medical term highlighting in transcript
• Welcome wizard (5-screen setup) and 2 demo scenarios (medical + dental)
• Subscription/licensing via Supabase + Stripe (14-day trial)
• 10 keyboard shortcuts
• Built with Tauri v2 for macOS (Apple Silicon)