Eighty-six points and 47 comments on Hacker News. Fifty-two stars on GitHub. Those two numbers sitting next to each other tell you almost everything about where SCM, the macOS photo and video search tool from developer allenv0, currently stands: plenty of people wanted to talk about it, and far fewer went and grabbed it.
That gap is not a knock on the project. It is the most interesting thing about it. A comment-to-star ratio like that usually means one of two things — either the idea provoked an argument, or the idea landed but the installation bar sat higher than the curiosity. With a tool that runs local AI vision models through Ollama, I would bet on the second. Possibly both.
What SCM actually claims to do
The pitch is deep AI search for every photo and every frame of video in any folder on macOS. Not your Photos library specifically, not a walled-off app database — any folder. The tool runs local vision models to analyze images, automatically renames files based on what it sees, and makes the whole collection searchable on your own hardware. It is MIT licensed. There are two forks.
Two parts of that deserve separate attention, because they are very different engineering problems wearing one product name.
Photo search by content is a solved-ish problem at this point. Vision models describe an image, you index the description, you search the index. Running it locally through Ollama instead of shipping your pictures to someone’s cloud is the actual selling point, not the search itself.
Every frame of video is the harder claim. Video is not a pile of photos — it is an enormous pile of photos, most of which are nearly identical to the one before. How a tool handles that determines whether it is useful or whether it quietly eats your afternoon and your battery. The facts I have do not tell me how SCM samples frames, deduplicates similar ones, or decides when a scene has changed. That is the first thing I would want to know before running it on anything larger than a test folder.
Why the local angle matters more than the feature list
Your photo library is the most personal dataset you own. It has your kids, your medical paperwork you photographed for an insurance claim, screenshots of your bank app, the whiteboard from a meeting you probably should not have photographed. Every cloud-based “search your photos with AI” product asks you to be okay with that leaving your machine.
SCM’s answer is that it does not. The analysis happens on your hardware through Ollama, and the index stays with you. For a certain kind of user — and judging by that Hacker News thread, a lot of AI practitioners are that kind of user — this is the whole reason to care. The feature is table stakes. The privacy model is the product.
The tradeoff is the one local AI always makes. You trade someone else’s data center for your own fan noise. Indexing a large library with a vision model is not a background task you forget about; it is a job you start and come back to. Nobody has published numbers I can point to here, which is exactly why I would treat any throughput claim with suspicion until someone benchmarks it on a real library.
The automatic renaming decision
This is the part I would push back on hardest, and it has nothing to do with the AI.
Renaming files based on what a model sees in them is a clever convenience and an irreversible one. If the model is right, your folder suddenly makes sense. If the model is wrong, you now have a thousand confidently mislabeled files and no original names to fall back on. Any tool that writes to filenames should make that step opt-in, reversible, and loudly explained before the first run. I have no information on whether SCM does any of that, and that is precisely the sort of thing I check before letting a new tool loose on a directory I care about.
Where I land
Fifty-two stars is early. Two forks is very early. MIT licensing and a public repo mean you can read the code before you trust it, which is more than most commercial photo search tools offer, and it means the project can be picked apart by the same crowd that showed up to comment.
My honest read: SCM is solving a real problem that the big platforms solve in a way a lot of people dislike. The local-first approach is the right instinct. The video frame search is the ambitious bet, and the renaming behavior is the thing I would want guardrails around. None of that is settled by a Show HN thread.
If you have a spare afternoon, a backed-up folder, and Ollama already running, this is worth a test. If you are looking at it as a drop-in replacement for however you currently find photos, give it a few more release cycles and a few more forks. Early tools reward careful users and punish eager ones.
🕒 Published: