Chop long YouTube videos into viral shorts without SaaS rent

What it does

Feed it a YouTube link or local file and it downloads the source, transcribes audio locally with faster-whisper, and classifies the content type. An LLM then scans the transcript for viral moments—emotional peaks, conflict, quotables—and scores candidates 0–100. The top N segments are auto-cropped into vertical 9:16 shorts, optionally prefixed with an AI-generated hook, with no watermarks or pre-roll credits.

The interesting bit

Instead of simple scene detection, the tool tries to algorithmize taste. The LLM evaluates hook moments, opinion bombs, and story peaks against a rubric, then deduplicates overlapping candidates so you do not get the same soundbite twice.

Key highlights

• Transcription runs fully offline via faster-whisper; only the LLM scoring and hook generation require an API key.
• Explicitly built as a free, watermark-free alternative to OpusClip and Vidyo.ai.
• Content-type aware: the LLM tunes its highlight prompt based on whether the input is a podcast, interview, tutorial, or vlog.
• Supports arbitrary aspect ratios and resolutions, not just the default 9:16.
• Ships with both a CLI and a lightweight local web UI for batch queuing.

Caveats

• Translation and voiceover are touted in the description but missing from the documented pipeline; the README only explains transcription, ranking, and cropping.
• The web interface is a thin wrapper around the same CLI logic—LLM providers and API keys still live in .env.
• You must supply your own LLM API key; there is no fully offline mode for highlight scoring.

Verdict

Ideal if you are sitting on hours of long-form content and want algorithm-ready shorts without paying SaaS rent. Skip it if you need a fully air-gapped workflow or frame-accurate editorial control.