Give Spotter a few reference photos of your thing and point it at your footage. It finds appearances of that thing, timestamps each to the half-second, and scores every one — with a sensitivity dial that runs from a few clean hits to everything. Finding runs entirely on your machine — no account, no upload, no per-video cost.
A scoreboard grades the three matchers against the cloud verdicts and your own thumbs, and your Agree/Disagree votes can retune the thresholds for the next scan.
Spotter is complete without the cloud. Flip the toggle and a vision model — Gemini, through OpenRouter, with your own key — gives a second opinion on a candidate crop. It sends the reference images and that one crop, never your video. The privacy envelope is enforced in code: US-only providers, zero retention, no data collection, and it fails closed rather than rerouting. Without a key, scanning and matching work fine.