LocalSnip Support & Help Center

LocalSnip uses on-device AI to transcribe your video's audio and then analyzes the transcript to identify moments with high viral potential. Each suggested clip receives a score and an explanation of why the AI selected it, so you can quickly evaluate which moments to keep. Transcription is powered by WhisperKit and clip analysis uses MLX — both run entirely on your Mac, optimized for Apple Silicon's GPU.

No. LocalSnip runs 100% offline after the initial AI model download. Your videos, audio, transcripts, and editing data never leave your Mac. The only network activity is downloading AI models from Hugging Face during setup or when you switch models. This makes LocalSnip ideal for client work and pre-release material.

LocalSnip supports five aspect ratios: 9:16 vertical (TikTok, Reels, Shorts), 16:9 widescreen, 1:1 square, 4:5 portrait, and the original source ratio. For captions, you can choose burned-in captions, closed captions (subtitle track), or both — plus SRT, VTT, and plain text transcript exports. The smart crop feature with rule-of-thirds guides helps keep your subject perfectly centered in any ratio.

LocalSnip requires a Mac with Apple Silicon (M1 or later) running macOS 14.0 Sonoma or later. The app uses the MLX framework, which is optimized for Apple Silicon's GPU, to run AI models locally. You can choose between compact models for speed or premium models for deeper analysis, depending on your hardware and preferences.

Yes. Once you've downloaded your preferred AI models during initial setup, LocalSnip works entirely offline. There are no server queues, no subscription data-mining, and no footage leaves your machine. An internet connection is only needed to download or update AI models.

LocalSnip gives you full creative control. Use the interactive timeline to review AI suggestions and drag boundaries to adjust them with precision. You can Option-drag to create entirely custom clips, or click anywhere in the transcript to jump to that point in the video. The app supports up to 50 levels of undo/redo so you can experiment freely.