Paste a URL. It runs the two on-device AI models StudentKeeper/ContentKeeper ships - the ONNX text classifier and the NSFW image model - plus its actual proxy ruleset, decrypted out of proxyDetection.dat with the extension's own wasm.
No server on this page, so the models run in the browser. Point it at the extension's models folder once - the files are cached here afterwards.
Wants models/captureClassification/model.onnx, models/model.json and models/group1-shard1of1. Running node server.js next to them skips all of this.
This reproduces the parts of the filter that run on the device: the ONNX text classifier used for safety categorisation, the NSFW image model with the extension's own thresholds, and its real proxy ruleset scored by a port of its own rule engine. Tick "open it in a real browser" and the eight rules that watch a running page - API hijacking, img-src rewriting, WebSocket traffic, in-page navigation - get judged for real too. What it still cannot reproduce is the server-side URL category database and your admin's policy. So "clear" does not mean "allowed", it means "nothing the local AI would flag from one fetch".