> Can the AI agent that generates the transformations be run locally?
We've tried. While building this, we tested just about every top model you can think of (closed and open). We weren't able to get any of the open models to perform to our standards. We have a background in model training/fine tuning, so I'm sure some further tuning could even the playing field but definitely getting a bit ahead of myself.
> What makes its results deterministic? Is it a "pick from a menu of curated transformations"?
When you send a request to generate a script (e.g. "change to dark mode"), the agent analyzes your current page (e.g. read through the CSS to look for relevant selectors) and generates targeted edits to achieve the desired results. The response is deterministic JS/CSS that can be applied on each page load (all local after the initial generation).
> What is the risk level of it generating something harmful? (Eg. That directly or inadvertently leaks data I don't want to leak)
We have a permissions framework inspired by Greasemonkey. For example, your script can only make web requests if it includes a grant for GM_xmlhttpRequest. That should clue you in if for some reason a given script is doing something unexpected. You can also dig into the options and find the scripts themselves if you want a thorough review