Guided VOD editing — AI maps every second of your stream into topics and beats (highlight, core, context, chat, transition, break) so you see the terrain before you cut.

VideoTool runs a 3-step pipeline on any stream VOD:

ingest → transcribe → detect narrative beats

1. Ingest — downloads the video (Twitch URL or local file), extracts audio
2. Transcribe — sends audio to Groq Whisper (fast, free tier: ~2h/day)
3. Detect beats — sends the full transcript to Claude in a single call, gets back topics + beats (highlight/core/context/chat/transition/break) with timestamps and confidence scores, tiling the entire stream

The output is a beats.json file that maps the narrative structure of your stream.

Paste a Twitch VOD link or drop a video file. The app runs the pipeline and shows an interactive beat timeline with click-to-seek video playback.

# Build and run the desktop app
cd src-tauri && cargo tauri dev

# Full beat detection pipeline — one command
videotool beats path/to/video.mp4
videotool beats https://twitch.tv/videos/<id>

# With language hint
videotool beats path/to/video.mp4 --language fr

# Re-run beat detection only (skip ingest + transcribe)
videotool llm-beats projects/<id>

# JSON progress output (for Tauri IPC)
videotool beats path/to/video.mp4 --json-progress

Share your beat timeline as a URL with an embedded Twitch VOD player:

# Upload a project's beats to the web viewer
videotool share ~/.videotool/projects/<id>

# Returns a shareable URL like:
# https://vodtool-api.<your-worker>.workers.dev/v/0123456789abcdef0123456789abcdef

The shared viewer shows the beat timeline synced to the Twitch VOD embed. Works for anyone with the link — no install needed. Links expire after 90 days.

Requires VITE_API_PROXY_URL and PROXY_AUTH_TOKEN in .env.

For paywalled VODs or local OBS recordings — no install required:

1. Open tools/beat-viewer.html in any browser
2. Load your video file + paste the beats.json output
3. Click beats to seek, hover for details

• Python 3.9+
• ffmpeg installed and in PATH
• yt-dlp installed and in PATH (for Twitch VOD downloads)
• GROQ_API_KEY — for transcription (free at console.groq.com)
• ANTHROPIC_API_KEY — for beat detection (Claude Sonnet)

git clone https://github.com/jbmoutout/videotool
cd videotool
python3 -m venv .venv && source .venv/bin/activate
pip install -e .

Copy .env.example to .env and fill in your API keys.

The beats.json schema:

{
  "topic_id": "coupe_cheveux",
  "topic_label": "Les 4 phases d'une coupe",
  "beats": [
    {
      "type": "context",
      "start_s": 1068.5,
      "end_s": 1185.2,
      "confidence": 0.75,
      "label": "Ça fait trois semaines, là c'est le pire stade de la repousse"
    },
    {
      "type": "core",
      "start_s": 1406.8,
      "end_s": 1767.3,
      "confidence": 0.90,
      "label": "Théorie des 4 phases d'une coupe de cheveux"
    },
    {
      "type": "transition",
      "start_s": 2066.4,
      "end_s": 2869.8,
      "confidence": 0.78,
      "label": "Débat sur les coupes longues, buzz cut et couvre-chefs de darons"
    }
  ]
}

Beat types:

• highlight — the most attention-grabbing moment (where a YouTube video should start)
• core — the meat of the topic, the highest-value segment
• context — setup and background that gives the core meaning
• chat — chat interaction, audience engagement
• transition — wind-down, pivot to the next topic
• break — off-topic pause, BRB, dead air

Not every topic has all 6 types. A short tangent might only have a highlight and core.

A Cloudflare Worker proxy (cloudflare-worker/) lets you distribute the app without requiring users to configure API keys. The proxy adds your keys and forwards requests.

cd cloudflare-worker

# Set up local dev
echo "GROQ_API_KEY=your-key" > .dev.vars
echo "ANTHROPIC_API_KEY=your-key" >> .dev.vars
echo "PROXY_AUTH_TOKEN=$(openssl rand -hex 16)" >> .dev.vars

# Run locally at http://localhost:8787
npx wrangler dev

# Deploy to production
npx wrangler deploy
wrangler secret put GROQ_API_KEY
wrangler secret put ANTHROPIC_API_KEY
wrangler secret put PROXY_AUTH_TOKEN

The original topic detection pipeline is still available for power users:

# 5-step pipeline: ingest → transcribe → chunks → embed → llm-topics
videotool pipeline path/to/video.mp4

# Individual steps
videotool transcribe projects/<id>
videotool chunks projects/<id>
videotool embed projects/<id>
videotool llm-topics projects/<id>
videotool list-topics projects/<id>
videotool show-topics projects/<id>
videotool cutplan projects/<id> --topic topic_0001
videotool export projects/<id>

Groq Whisper free tier covers ~2 hours of audio per day.

MIT