SCM Brings Private, Local AI Search to Photos and Video on macOS

Summary

SCM is a macOS application for searching photos and videos with AI while keeping media on the user’s computer. Its Files mode uses switchable ONNX Runtime vision models to rank whole images and videos by natural-language meaning, with filename and phrase boosts plus explanations for why each result matched. A Scenes mode segments videos with ffmpeg, embeds sampled midpoint frames, and opens results at the matching timecode; whole-video results use three sampled frames. Separate OCR and Dialogue modes provide literal text retrieval from images or video frames through Tesseract and exact spoken-word search through Whisper, including highlighted matches and direct seeking. Users can optionally enable a local llama.cpp chat sidecar using extracted dialogue, OCR, and filename evidence, with numbered citations and no model execution until the feature is turned on. The app watches folders, imports and content-hash deduplicates files, retries problem files, and can re-embed a library in the background when the active model changes. It offers four vision models, two English Whisper sizes, configurable video sampling presets, and downloadable language packs; the first default vision model download is about 435 MB, after which processing is offline. Search tabs, screenshot and email views, saved queries, embedding snapshots, and macOS settings organize the library. Installation targets Apple Silicon Macs running macOS 12 or later through a Homebrew cask, with Bun-based development and Electron packaging. The project states that it has no accounts, telemetry, uploads, or cloud inference, while weights and optional components are downloaded once and verified before local use.