Your three-hour podcast, cropped into TikTok bait — locally

What it does

You feed it a long video or audio recording. It extracts audio, transcribes it (Deepgram in the cloud, or Whisper locally), asks an LLM to find the moments most likely to perform as short-form content, and center-crops the source into vertical H.264 clips via native ffmpeg. Everything — transcripts, clip candidates, project data — lives in a local SQLite database, managed through a native project dashboard.

The interesting bit

The “local-first” claim is real but honest about its limits: you can run fully offline with Ollama models plus openai-whisper, yet the README itself warns that 3B/7B local models are bad at the actual job — they can’t reason over long transcripts or compute accurate segment timestamps, and tend to emit fragments that are too short. The recommended path is DeepSeek, which the author prices at under $0.001 per transcript. So: local storage, cheap cloud brains.

Key highlights

• Tauri 2 + React + Rust + SQLite desktop app for macOS, Windows, and Linux
• Multi-LLM moment detection: DeepSeek (default), Claude, or local Ollama models
• Fully offline mode available via Ollama + Whisper transcription
• Automated pipeline: import → transcribe → rank → portrait-crop, one chain
• Built-in UI warnings for missing environment variables and unsigned-install friction

Caveats

• Requires FFmpeg/FFprobe on your PATH; macOS users may need a custom Homebrew tap for full caption rendering support
• Local LLM mode exists but the README explicitly doesn’t recommend it for moment detection
• Builds are self-signed, so expect Gatekeeper and SmartScreen warnings on first launch

Verdict

Worth a look if you produce podcasts or streams and want a cheap, privacy-respecting clip-mining pipeline instead of a $30/month web tool. Skip it if you need polished auto-edited captions out of the box or refuse to touch an API key.