scribe-minutes 0.1.0

Local, privacy-preserving meeting transcription and minutes generation for macOS. Parakeet TDT for speech-to-text, any OpenAI-compatible LLM for minutes.

Scribe

Local, privacy-preserving meeting transcription and minutes for macOS. Nothing leaves your machine.

Drop an audio or video file into the app, get a transcript and structured meeting minutes back. Record directly in the app if you prefer. Everything runs on your hardware — no cloud APIs, no data exfiltration, no subscriptions.

What It Does

1. Speech-to-text via Parakeet TDT — Nvidia's speech model running on Apple MLX (GPU-accelerated on Apple Silicon). ~600MB model, processes audio in 2-minute chunks.
2. Meeting minutes via any OpenAI-compatible LLM — takes the transcript and produces attendees, agenda, decisions, action items. Works with local models (Qwen, Llama, Mistral via mlx-lm/ollama/vLLM) or remote APIs.
3. MCP server for LLM tools — Claude Code, Gemini CLI, or any MCP client can transcribe files and generate minutes directly.

Install

Requirements: macOS 15+ on Apple Silicon, Python 3.12+, ffmpeg.

# Install ffmpeg if you don't have it
brew install ffmpeg

# Install the backend
pip install scribe-minutes
# or: uv pip install scribe-minutes

# Start it
scribe
# → Backend running at http://localhost:8890

From source

git clone https://github.com/aaronmontgomery/scribe.git
cd scribe
pip install -e .   # or: uv pip install -e .
scribe

macOS App

The native SwiftUI app provides drag-drop, recording, and a two-pane results view. Requires xcodegen.

brew install xcodegen
make app-install    # builds and copies to /Applications

Or build manually:

cd app
xcodegen generate
xcodebuild -project Scribe.xcodeproj -scheme Scribe_macOS -destination 'platform=macOS' build

Configuration

All configuration is via environment variables. None are required — the backend works out of the box for transcription. Minutes generation needs an LLM.

Variable	Default	Description
SCRIBE_LLM_URL	http://127.0.0.1:18080/v1/chat/completions	OpenAI-compatible chat endpoint for minutes
SCRIBE_LLM_MODEL	mlx-community/Qwen3.5-35B-A3B-4bit	Model name to request
SCRIBE_LLM_HOST	(empty)	SSH host for auto-tunnel to remote LLM (e.g. my-server)
SCRIBE_LLM_REMOTE_PORT	8080	LLM port on remote host
SCRIBE_LLM_LOCAL_PORT	18080	Local port for SSH tunnel
SCRIBE_LLM_TIMEOUT	300	LLM request timeout (seconds)
SCRIBE_RAG_URL	(empty)	Optional RAG MCP endpoint for name/role correction
SCRIBE_DEVONTHINK_GROUP	(empty)	DEVONthink group UUID for export (saves to inbox if unset)

Example setups

Local LLM via Ollama:

export SCRIBE_LLM_URL=http://127.0.0.1:11434/v1/chat/completions
export SCRIBE_LLM_MODEL=qwen2.5:32b
scribe

Remote LLM via SSH tunnel:

export SCRIBE_LLM_HOST=my-gpu-server
export SCRIBE_LLM_REMOTE_PORT=8080
scribe
# Backend auto-creates: ssh -N -L 18080:127.0.0.1:8080 my-gpu-server

Transcription only (no LLM needed):

scribe
# Just use "Transcript Only" mode in the app, or the /transcribe API endpoint

MCP Integration

LLM tools (Claude Code, etc.) can transcribe files directly.

Add to your MCP config:

{
  "scribe": {
    "type": "http",
    "url": "http://127.0.0.1:8890/mcp"
  }
}

Tools:

• transcribe_file(path, format) — speech-to-text (txt, srt, vtt, json)
• generate_minutes(path) — transcript + structured minutes
• batch_minutes(paths) — process multiple files
• get_latest_output() — read the most recent result

API

Endpoint	Method	Description
/health	GET	Backend status
/transcribe	POST	Upload file → transcript
/transcribe/path?path=...	POST	Local file → transcript
/minutes	POST	Upload file → transcript + minutes
/minutes/path?path=...	POST	Local file → transcript + minutes
/batch/minutes	POST	Multiple files → batch results
/output/list	GET	List saved outputs
/output/latest	GET	Latest output as markdown
/export/devonthink	POST	Save to DEVONthink (macOS)

Architecture

┌──────────────────────────────────┐
│  SwiftUI App                     │  Native macOS: drag-drop, recording,
│  Record / Open / Batch           │  Liquid Glass UI, keyboard shortcuts
└──────────┬───────────────────────┘
           │ localhost:8890
┌──────────▼───────────────────────┐
│  Python FastAPI Backend          │
│  ├─ Parakeet TDT (STT on MLX)   │  ~600MB, runs on Apple GPU
│  ├─ ffmpeg (audio extraction)    │  video → 16kHz mono WAV
│  ├─ LLM client (minutes)        │  any OpenAI-compatible endpoint
│  └─ /mcp endpoint                │  MCP server for LLM tools
└──────────────────────────────────┘

Three ways to use the same backend:

1. SwiftUI app — drag-drop GUI
2. MCP tools — Claude/Gemini/Codex transcribe via tool calls
3. REST API — curl, scripts, integrations

macOS App Features

• Drag-drop or Cmd+O for single files, Cmd+Shift+O for batch
• Built-in audio recorder (click Record, speak, get minutes)
• Liquid Glass UI (macOS 26+)
• Two-pane results: transcript left, rendered markdown minutes right
• Keyboard shortcuts: Cmd+S (save), Cmd+E (DEVONthink), Cmd+Shift+C (copy minutes), Cmd+N (new)
• Drag transcript/minutes directly into other apps
• Recent transcriptions on the home screen
• System notification when processing completes
• DEVONthink export (macOS)

License

MIT