An open-source OpusClip built from Whisper and prompts

What it does — It’s a Python pipeline that ingests YouTube videos or local files, transcribes audio locally with faster-whisper, and uses a remote LLM to identify the most viral segments. It then auto-crops those highlights into vertical 9:16 shorts, optionally prepending an AI-generated hook and burning in subtitles. The goal is a self-hosted, watermark-free alternative to OpusClip or Vidyo.ai.

The interesting bit — The LLM doesn’t just blindly guess highlights. It first classifies the video genre and pacing—podcast, interview, tutorial, vlog—then scores transcript segments against an explicit virality framework: hook moments, emotional peaks, “opinion bombs,” revelations, conflict, and practical value. That structured prompt engineering is doing the analytical work; the rest of the pipeline is largely orchestration around Whisper and the video renderer.

Key highlights —

• Local transcription with faster-whisper; only the virality ranking requires a paid LLM API key (OpenAI, Gemini, or MuAPI)
• Content-aware prompting adjusts highlight detection per video type rather than using a one-size-fits-all heuristic
• Optional AI-generated hook can be disabled with --no-hook
• Ships with both a CLI and a minimal local web UI for queuing multiple videos
• Output clips carry no watermarks or pre-roll credits

Caveats —

• The README advertises translation and voiceover in the tagline, but the detailed pipeline only covers transcription, highlight ranking, and cropping—those extra features are unclear
• You still need to fund your own LLM API usage; “free” refers to the codebase, not the inference
• The web interface is functional but bare-bones, with a split backend/frontend setup that feels more like a local prototype than a packaged app

Verdict — Ideal for creators who want algorithmic short-form reels without SaaS subscriptions, but anyone needing precise manual editorial control will find this is still very much prompt-and-pray automation.