Local AI search for every photo and every frame of video on macOS.
Type bride walking down aisle — get the exact clip. Nothing ever leaves your Mac.
Free CLI for macOS. No upload, no account, no cloud. Pro is $29 one-time.
A single wedding shoot produces 200 GB of footage across three cameras and four hours of continuous recording. Finding the ninety seconds where the bride walks down the aisle currently means scrubbing timelines by hand, or trusting a filename like CAM_A_0047.MOV that tells you nothing about what is inside. The same problem scales to family archives with 80,000 photos and to podcast studios sitting on years of unindexed b-roll. The footage exists. The ability to find it does not.
Existing tools solve the wrong half of the problem. Apple Photos searches your library beautifully, but it refuses to index external drives and folders — exactly where professionals keep their masters. Adobe Premiere and DaVinci Resolve offer scene detection and speech search, but only after you import everything into a project, and only inside a $22.99/month subscription. Cloud semantic search services want you to upload gigabytes of client footage to a server you do not control, which most working creators will never do. The gap is obvious: local, frame-level, folder-native search that treats a directory of raw files as a first-class library.
The timing is what makes this buildable today. Open CLIP models run comfortably on Apple silicon, MLX and CoreML give native acceleration without a GPU server, and ffmpeg makes frame extraction a solved problem. A developer discussion thread on this exact topic reached 163 points and 73 comments in a single day, with commenters repeatedly asking "can it search X?" — a live, unfiltered list of the queries real users want to run. That demand signal, combined with zero-upload inference, means the first tool that ships a credible local index owns the category before the incumbents notice.
Give FrameSeek a path like ~/Movies/wedding_2026/. It walks every video
and photo file, extracts one frame every two seconds with ffmpeg, and builds a local
vector index. A four-hour shoot becomes roughly 7,200 frames — a few minutes of
one-time work on an M-series Mac.
Describe the moment in plain language: "red suitcase on the platform", "bride walking down aisle", "sunset over the water". CLIP encodes your query into the same vector space as the frames, so matching is a similarity search rather than a keyword lookup. No tags, no metadata, no manual labeling required.