SCM searches every photo and every video frame on a Mac from a plain-language description, offline

Finding a specific moment in years of photos and videos usually means scrolling by date, because filename and keyword search cannot see what is in the picture. SCM, short for Screen Memories, is a new open-source macOS app that lets you type what you remember in plain language and ranks every photo and every video frame in a folder by how well it matches, with all of the inference running on the local machine.

The repository at allenv0/SCM was published on 2026-10-03 and gained 84 stars in one day. A Show HN post titled "AI search for every photo and every frame of video on macOS" has 42 points on Hacker News at the time of writing. The project is packaged as an Electron app and installs through a Homebrew cask on Apple Silicon machines running macOS 12 or later.

Five ways to search the same library

The README lists five search modes a single query can run against. Files ranks whole photos and videos by meaning, using a vision model to score cosine similarity against image embeddings. Scenes segments every video into shots, embeds the midpoint frame of each one, and jumps to the exact timecode when a scene is picked. OCR matches text visible inside images and frames, using Tesseract with English on by default and 35 other languages toggleable. Dialogue retrieves the exact spoken words from Whisper transcripts of videos, with three tiers of exactness. An opt-in LLMs mode runs a small local chat model over the dialogue, OCR text and filenames the app has already extracted, with citations the user can click.

The vision model is CLIP ViT-L/14@336 by default, with three SigLIP variants switchable from Settings. Each one downloads once from Hugging Face, around 435MB for the default, and the author says switching models re-embeds the whole library in the background without blocking search. Whisper weights are either tiny.en at about 150MB or base.en at about 300MB, downloaded on first use.

Everything stays on the machine

The author describes the design as local-first, with no accounts, no cloud and no uploads, and media copied into an app-managed library under ~/Library/Application Support/scm and streamed from disk. The renderer runs in an Electron sandbox pinned to app:// with a Content Security Policy locked to self. Only main-process workers ever download anything, and the README says each download is SHA-256-verified. There is no telemetry.

The practical consequence is that a photo library with family pictures, personal screenshots or sensitive documents never leaves the computer. The project does require an internet connection the first time each model is used, to pull the weights, and after that the author states that everything is offline. The current version is 0.2.4 and the project is Electron plus Bun plus React.

The repository is young, with 84 stars in 24 hours and no release-notes history to judge long-term pace. Anyone trying it is trying it early, and the models and the OCR language packs each add hundreds of megabytes to disk. For a reader who has given up on finding a specific photo through a search box, the trade is a few gigabytes of model weights against being able to type "the picture from the lake in winter" and get the picture.

Source

Primary source: allenv0/SCM on GitHub. Discussion on Hacker News.