Remember when the only way to find one specific photo was scrolling. Thumb on the screen, eyes glazed, hunting for a receipt you photographed in March or the whiteboard from that meeting where someone actually had a good idea. You knew the picture existed. You just had no way to ask for it.
Text got solved years ago. You can search a folder of documents for a single word and get an answer before you finish typing. Images never got that treatment, because a file named IMG_4471.HEIC tells your computer absolutely nothing about what’s inside it. The pixels are there. The meaning isn’t.
A small macOS project called SCM, built by a developer going by allenv0, takes a swing at that gap. It does AI search across every photo and every frame of video on your Mac. Not every file, every frame. That distinction matters more than it sounds like it does, and I’ll get to why.
What it actually does
SCM runs local AI vision models over your images, figures out what’s in them, and renames the files based on what it sees. The result is a library you can search the same way you’d search your email. The processing happens on your own hardware through Ollama, which is a tool for running AI models locally instead of sending data to someone else’s servers.
So the flow looks roughly like this:
- A vision model looks at each photo and identifies what’s in the frame
- The file gets renamed to reflect that content
- Your existing search tools can now find things by description instead of by filename gibberish
- Nothing leaves your machine
That last point is the one I’d underline for anyone who’s been nervous about handing a decade of family photos to a cloud service. Your photo library is one of the most personal datasets you own. It has your kids, your medical paperwork, your passport, the inside of your home. The privacy math changes completely when the analysis runs locally.
Why the video part is the interesting bit
Searching photos is a solved-ish problem in the sense that several large companies offer some version of it. Video is messier. A two-minute clip is thousands of individual images stitched together, and most search tools treat the whole thing as one object with one label. You can find the video. You can’t find the moment.
Frame-level search flips that. If the tool is looking at every frame, then the thing you remember from a clip becomes findable, even if it was on screen for two seconds. For anyone who records lectures, documents work sites, keeps footage of a project in progress, or just has four years of phone videos nobody has ever rewatched, that’s a different kind of useful.
The reception, in context
SCM showed up as a Show HN post on Hacker News last week and pulled 86 points with 47 comments. The GitHub repository sits at 52 stars with two forks, under an MIT license.
Those are small numbers, and I want to be honest about that rather than dress them up. This is not a product launch. It’s an early project getting a look from a technical crowd. But the comment-to-point ratio is the part I find telling. Forty-seven comments against 86 points means people weren’t just upvoting and moving on, they were arguing, asking, and poking at it. On Hacker News, that pattern usually means a project touched a real problem rather than a novelty.
What this tells us about where AI tools are heading
If you’re trying to understand AI agents without a computer science degree, SCM is a clean example of a pattern that keeps repeating. A capable model does one narrow job repeatedly, on your behalf, without you supervising each step. You don’t prompt it a thousand times. You point it at a folder and it works through the pile.
That’s the practical shape most of this technology takes once the demos stop. Not a chatbot you converse with. A quiet process that handles the tedious part of something you already wanted to do.
The local-first angle is the other signal. Running models on your own machine used to mean accepting worse results. Tools like Ollama have made it ordinary enough that a solo developer can build a photo search tool on top of it and ship it under an open license. That’s a meaningful shift in who gets to build this stuff.
Should you try it
If you’re on macOS, comfortable installing developer tools, and have a photo or video folder you’ve given up on organizing, it’s worth a look. Fifty-two stars means you’re an early tester, not a customer, so expect rough edges and read the repository notes before you point it at anything irreplaceable.
And if you’re not technical, treat this as a preview. The idea that your own files should be searchable by meaning, privately, on hardware you own, is not going to stay a niche project for long.
🕒 Published: