RoughCut

RoughCut is a local-first rough cut generator for Final Cut Pro on macOS. Point it at a talking-head video and it will remove long silences and optionally remove bad takes and fluff, producing FCPXML variants that relink to the original media.

What it does

• Removes long silences using FFmpeg silencedetect.
• Tightens low-energy sections with an RMS energy gate.
• Optionally removes bad takes using a local disfluency model + retake detection.
• Optionally removes fluff/off-topic content using Gemini.
• Produces a pack of .fcpxml variants for import into Final Cut Pro.

Quickstart (Local)

One command (auto-installs dependencies on first run):

bash scripts/dev.sh

• Opens the web UI at http://localhost:5173.
• Starts the API server at http://127.0.0.1:8000.

API server only (headless):

bash scripts/dev.sh start-server

Stop whatever is running:

bash scripts/dev.sh stop

Use With OpenClaw

See docs/openclaw.md.

Headless runner (what agents execute):

bash scripts/openclaw/roughcut.sh \
  --video "/absolute/path/to/video.mp4" \
  --out "/absolute/path/to/output_dir" \
  --remove-bad-takes true \
  --remove-fluff false

If you want RoughCut to download the input video first (recommended for agent-hosted Macs), pass a direct HTTPS download URL:

bash scripts/openclaw/roughcut.sh \
  --video-url "https://example.com/path/to/video.mp4" \
  --out "/absolute/path/to/output_dir" \
  --remove-bad-takes true \
  --remove-fluff false

Outputs

Each job writes to server/jobs/<jobId>/ and produces:

• xml_variants.zip (on disk)
• RoughCut.xml_variants.zip (download filename)
• VIDEO_STEM.<variant>.fcpxml files inside the zip
• debug.zip (on failures)
• transcripts and decision JSONs for debugging

Import into Final Cut Pro:

1. Unzip RoughCut.xml_variants.zip.
2. In Final Cut Pro: File -> Import -> XML...
3. Choose the .fcpxml variant you want.

Configuration

• Runtime defaults come from .env (see .env.example) and server/app/settings.py.
• UI settings are stored in server/config.json (git-ignored) and override .env.

LLM keys

• GEMINI_API_KEY is only required if you enable fluff removal.
• OPENAI_API_KEY is optional (if you switch provider to OpenAI in the UI).

Troubleshooting

• First run may build whisper.cpp and download models. Ensure:
  • Xcode Command Line Tools: xcode-select --install
  • CMake: brew install cmake
• Missing ffmpeg: install via Homebrew or place binaries in ./bin and add to PATH.
• If ports are stuck (5173/8000): bash scripts/dev.sh stop
• Logs:
  • server/run/web.log
  • server/run/server.log
• whisper.cpp build failed on macOS: run xcode-select --install, then retry.
• dyld libwhisper.1.dylib missing: delete server/bin/whispercpp and restart to rebuild.
• whisper.cpp download failed: set WHISPER_CPP_PATH and WHISPER_CPP_MODEL_PATH to local paths, or set WHISPER_CPP_AUTO_DOWNLOAD=false and manage installs manually.
• Transcription is slow: ensure you are using a Metal-enabled whisper.cpp binary on Apple Silicon.

Security and privacy

• No media is uploaded to the cloud.
• Only transcript text and timestamps are sent to Gemini (fluff removal and/or XML variants), depending on your config.

Contributing

See CONTRIBUTING.md.

License

MIT (see LICENSE).