An Open-Source Viral-Clip Factory for Your YouTube Backlog

What it does Paste a YouTube link—or a local file path—and the tool downloads the video, transcribes it locally with faster-whisper, then uses an LLM to classify the content style and rank moments by virality. It renders the top candidates as vertical 9:16 clips, optionally prepending an AI-generated hook, with no forced credits or watermarks.

The interesting bit Instead of naive timestamp splitting, the LLM scores transcript segments through a specific virality framework—hook moments, emotional peaks, opinion bombs, conflict, quotables—and collapses overlapping candidates before rendering. It is essentially an automated editor with a taste for drama.

Key highlights

• Runs transcription entirely offline; only the highlight-ranking LLM hits the cloud (OpenAI, Gemini, or MuAPI).
• Content-type detection tunes the prompt per genre—podcast, interview, tutorial, vlog—before hunting for peaks.
• Optional AI-generated hook injection at the start of each clip unless --no-hook is set.
• Both CLI and a bare-bones local web UI for batch queuing.
• Outputs are clean: no pre-clip credits or watermarks.

Caveats

• Translation and voiceover are promised in the description, yet the README only details highlight detection and cropping; how those features operate is unclear.
• Virality scoring depends on a paid or rate-limited third-party LLM API; there is no fully offline path.
• The web interface is a manually served local page—functional, but do not expect Vercel polish.

Verdict YouTubers and podcasters sitting on long-form archives who want SaaS-free shorts should look here. If you need a fully offline pipeline or a polished hosted UI, this is not it yet.