Turn YouTube rambles into TikToks with a local AI virality judge

What it does

Paste a YouTube link—or point it at a local file—and the tool returns cropped 9:16 shorts ready to post. It downloads the source, transcribes it locally with faster-whisper, then hands the transcript to an LLM. The model classifies the content style—podcast, tutorial, vlog—and hunts for “viral” moments using a rubric of emotional peaks, conflict, and quotable lines, scoring each candidate 0–100. The top picks are rendered as vertical clips with optional AI-generated hooks and subtitles.

The interesting bit

Instead of naive scene detection, the LLM acts like a talent-show judge, explicitly scoring moments for hook potential, opinion bombs, and story peaks. It then deduplicates overlapping highlights before rendering, so you don’t get three versions of the same sentence.

Key highlights

• Transcription stays entirely local; only the LLM ranking hits the API.
• Supports OpenAI, Gemini, or MuAPI as the judge, with Gemini offering a limited free tier.
• Optional AI-generated hook injection at the start of each clip.
• Web UI included for queuing batches without touching the CLI.
• Outputs are watermark-free and skip pre-clip credit screens.

Caveats

• Requires an LLM API key; there is no fully offline mode.
• The README promises translation and voiceover in the tagline, but the documented workflow only covers transcription, highlight scoring, and cropping—how those extras work is unclear.
• The web UI is functional but minimal: it queues videos and sets flags, while API keys still live in a local .env file.

Verdict

A solid weekend project for creators or developers building content pipelines who’d rather spend API credits than SaaS rent. If you need a polished, fully offline, or hands-free experience, keep looking.