Scripta

Privacy-first meeting transcription & AI summary for macOS

Features • Screenshots • Installation • Requirements • Build • Contributing • License

▶ 90-second walkthrough: dual-channel capture → Gemma 4 summary with full 128K context

---

Scripta is a native macOS app that captures both your microphone and system audio during meetings, transcribes them in real-time, and generates AI-powered summaries — all running 100% locally on your Mac. No cloud. No subscriptions. No data leaves your machine.

• Mic transcription powered by whisper.cpp (Metal-accelerated, on-device)
• System audio transcription via Apple SFSpeechRecognizer (on-device)
• AI summaries & chat via Ollama (local LLM)

Features

Dual-Channel Real-time Transcription

• Mic ("You") via whisper.cpp — 100% local, Metal-accelerated on Apple Silicon
• System audio ("Remote") via Apple on-device speech recognition
• Supports 10+ languages (English, Chinese, Japanese, Korean, Spanish, French, German, and more)
• Mic mute button — instantly pause/resume mic recording during a session

AI-Powered Summaries

• Local AI summaries powered by Ollama
• Key points + action items extraction
• Streaming responses displayed in real-time

AI Chat Panel

• Ask questions about your meeting transcript
• Multi-turn conversation with context
• Works during recording or on completed sessions

Live Translation (macOS 15+)

• Real-time bilingual transcript using Apple Translation framework
• Context-aware translation for improved quality

Meeting History

• Browse and search all past sessions
• View transcripts, summaries, and re-generate with AI
• Audio recordings saved alongside transcripts
• Sessions stored in ~/Documents/ScriptaScripts/

Two Display Modes

• Full mode: complete UI with transcript, summary, chat, and controls
• Minimal mode: floating live captions bar — stays on top while you work

Privacy by Design

• All processing happens on your Mac
• No internet connection required (except for initial model downloads)
• No account, no telemetry, no tracking

Screenshots

Full Mode — Live Recording with Translation

Live dual-channel transcription with bilingual translation. The "Remote" channel captures system audio from any meeting app (Teams, Zoom, Meet, etc.) while "You" captures your microphone via whisper.cpp.

AI Summary & Chat

After recording, generate an AI summary with key points and action items using a local Ollama model. The Ask AI chat panel lets you query your transcript in natural language.

AI Model Setup — Pick your local LLM

Choose between Qwen 2.5 (fast, 1.8 GB) for everyday meetings or Gemma 4 (high quality, 7.2 GB) for hour-long sessions. Each row surfaces the model's true context window — Scripta now passes that value to Ollama via num_ctx, so a 60-minute meeting actually gets summarized end-to-end rather than the last five minutes. Recommended models can be pulled with one click.

Minimal Mode — Floating Captions

A compact floating caption bar that stays on top of your other windows — perfect for following along during a meeting without switching apps.

Meeting History

Browse all past sessions with duration, entry count, and audio availability. Search transcripts and re-generate summaries at any time.

Installation

One-Line Install (recommended)

curl -fsSL https://raw.githubusercontent.com/thehwang/Scripta/main/scripts/install.sh | bash

This single command will:

• Download the latest release for your macOS version
• Install Scripta.app to /Applications
• Install Ollama via Homebrew (if not already installed)
• Start Ollama as a background service
• Pull the default AI model (qwen2.5:3b)
• Download the Whisper speech model (ggml-base.bin, ~142 MB)
• Launch Scripta

One-Line Install — with Gemma 4 (recommended for long meetings)

curl -fsSL https://raw.githubusercontent.com/thehwang/Scripta/main/scripts/install.sh | SCRIPTA_INSTALL_GEMMA4=1 bash

Same as above but also pulls Gemma 4 E2B (~7.2 GB extra). Recommended if your meetings frequently exceed 30 minutes — Gemma 4's 128K context window summarizes the full transcript in one pass with no chunking. The default qwen2.5:3b is still installed alongside, so you can switch between them anytime in Settings → AI Model.

Manual Install

1. Download the latest .zip from Releases
2. Extract and move Scripta.app to /Applications
3. Clear quarantine: xattr -cr /Applications/Scripta.app
4. Install Ollama: brew install ollama && brew services start ollama
5. Pull a model: ollama pull qwen2.5:3b
6. Launch Scripta — it will prompt to download the Whisper model on first run

Requirements

Component	Requirement
macOS	14.0 (Sonoma) or later
Chip	Apple Silicon (M1/M2/M3/M4) or Intel
Ollama	Required for AI summaries and chat
Disk	~2 GB for AI model + ~142 MB for Whisper model
Translation	macOS 15+ (Sequoia) for live translation

Building from Source

git clone https://github.com/thehwang/Scripta.git
cd Scripta
make run

This will build whisper.cpp as a static library, build the app with Swift Package Manager, sign it with an ad-hoc certificate, and launch it.

Project Structure

Scripta/
├── Sources/
│   ├── Scripta/              # Main app
│   │   ├── AppDelegate.swift
│   │   ├── ContentView.swift        # Main UI (full + minimal modes)
│   │   ├── MeetingRecorder.swift     # Recording orchestration
│   │   ├── WhisperEngine.swift       # whisper.cpp integration for mic ASR
│   │   ├── SystemAudioCapture.swift  # ScreenCaptureKit system audio
│   │   ├── SummaryService.swift      # Ollama AI summary
│   │   ├── ChatPanel.swift           # AI Q&A sidebar
│   │   ├── HistoryPanel.swift        # Meeting history browser
│   │   ├── TranslationService.swift  # Apple Translation wrapper
│   │   ├── MeetingStore.swift        # Session persistence
│   │   └── ...
│   ├── ScriptaCore/          # Shared types (TranscriptEntry, logging)
│   └── CWhisper/             # whisper.cpp C bridging (systemLibrary)
├── benchmarks/               # Reproducible model + context-window benchmarks
│   ├── synthetic-transcript.md  # Fixture: 60-min fictional all-hands meeting
│   └── findings.md           # Qualitative analysis of each model's output
├── demo/                     # Hammerspoon automation that drives the UI
│   ├── scripta-demo.lua      # through the v3.2 launch demo (timing-synced
│   └── README.md             # to a pre-recorded voiceover for repeatable takes)
├── screenshots/              # App screenshots for README
├── scripts/
│   ├── install.sh            # One-line installer
│   └── benchmark_models.sh   # Run any installed Ollama model against fixture
├── Resources/
│   └── AppIcon.icns
├── Makefile                  # Build, install, deploy targets
└── Package.swift

Permissions

Scripta requires the following macOS permissions (prompted on first launch):

• Microphone — to capture your voice
• Screen Recording — to capture system/meeting audio via ScreenCaptureKit
• Speech Recognition — for on-device transcription

Legal Notice

Recording Disclaimer: Recording conversations may be subject to local, state, or national consent laws. Many jurisdictions (including parts of the United States, Canada, and the European Union) require all participants to be informed and to consent before a conversation is recorded. You are solely responsible for complying with all applicable laws when using this software. Scripta does not record by default — recording is explicitly initiated by the user. The developers of Scripta assume no liability for misuse or non-compliance with recording consent laws.

Privacy: Scripta processes all audio and transcription data entirely on your device. No audio, transcript, or personal data is transmitted to any external server. AI summaries are generated locally via Ollama. The only network activity is the one-time download of AI and speech models.

License

MIT — free for personal and commercial use.

---

Made with ❤️ on a Mac