An open-source YouTube-to-shorts tool that lets you audit why each clip was picked

Paste a YouTube URL, get a vertical short ready to post. That is the entire promise, and the implementation is more considered than the pitch suggests.

The transcription step runs locally via faster-whisper regardless of which LLM you configure — OpenAI, Gemini, or MuAPI. That separation matters: your audio never leaves your machine during the most compute-heavy part of the job. The LLM only sees the transcript, which it uses first to classify the video's content type (podcast, interview, tutorial, vlog, and so on) and then to rank candidate highlights against a virality framework the README spells out explicitly: hook moments, emotional peaks, opinion bombs, revelations, conflict, quotables, story peaks, practical value. Each candidate gets a score from 0 to 100. Overlapping segments are collapsed by score before the top N are selected, so you don't get three clips from the same thirty-second stretch.

The output is a 9:16 crop with optional AI-generated hooks prepended to each clip. You can suppress the hooks with --no-hook if you find them intrusive. Resolution goes up to 1080p; the default is 720. You can also feed a local file path instead of a URL, which is useful if you're working with footage that isn't on YouTube.

Gemini is the only provider with a free tier, though it carries a daily limit. OpenAI requires payment. MuAPI is pay-per-use without a subscription, which the README positions as a middle option. The Whisper model size is configurable from tiny through large-v3, and device selection (cpu, cuda, auto) is explicit in the .env.

There's also a local web interface — a separate server process plus a static frontend — for queuing multiple videos without touching the CLI. It doesn't change the LLM configuration; that still lives in .env.

The honest comparison the README makes is to OpusClip and Vidyo.ai. Those are subscription SaaS products; this runs on your hardware with your API keys and produces clips with no watermarks. The tradeoff is setup friction and the cost of whatever LLM calls you make. For anyone already paying for an OpenAI or Gemini key, the marginal cost per video is small. For someone starting from zero, Gemini's free tier is the path of least resistance, rate limits notwithstanding.

The virality scoring framework is the most interesting design choice here. Rather than letting the LLM freestyle on what makes a good clip, the README enumerates the specific signal categories it's prompted to look for. That makes the selection logic auditable in a way that black-box SaaS tools are not.