Opus Clip's worst nightmare is a Python script

What it does Feed it a YouTube URL (or local file) and it spits out ranked, auto-cropped 9:16 clips ready for TikTok, Reels, or Shorts. Under the hood: Whisper transcription, an LLM scoring moments by “hook moments, emotional peaks, opinion bombs, revelation moments, conflict, quotable lines, story peaks, and practical value,” then vertical reframing via either a hosted API or local OpenCV face-tracking.

The interesting bit The virality framework is fully exposed and editable in highlights.py — not a black-box SaaS algorithm. You can tweak what counts as viral, swap between OpenAI and Gemini for ranking, and even run mostly offline (--mode local) with only the LLM call leaving your machine. It also auto-chunks videos over 30 minutes with overlap so long-form content doesn’t get shortchanged.

Key highlights

• Dual-mode architecture: fast API path via MuAPI, or local path with yt-dlp + faster-whisper + ffmpeg
• Each clip gets a 0-100 viral score, a hook sentence, and a one-line explanation of why it should perform
• Smart deduplication collapses overlapping highlights by score
• JSON output (--output-json) for full pipeline automation
• MIT licensed, importable as a Python library (generate_shorts(...))

Caveats

• Default API mode leans on MuAPI, a commercial service the repo promotes fairly heavily (it’s also the project’s apparent business model)
• Local mode still requires an LLM API key; there’s no fully offline option for highlight ranking
• The README mentions gpt-5-mini for API-mode LLM ranking, which doesn’t exist — likely a typo for GPT-4o mini or similar

Verdict Worth a look if you’re a creator, agency, or developer who wants clip generation without subscription lock-in and doesn’t mind some API key juggling. Skip it if you need a fully offline, zero-API solution or if you trust your own taste more than an LLM’s idea of “opinion bombs.”