Ask HN: Practices for sandboxing LLM agent front ends and IDEs?

I've been using Cursor for a while, and I've increasingly had some unease about it having unfettered read access to my home directory. Contemporaneously, I'm starting another round of T&E with the various LLM options, starting with Claude Code. I gave it a quick whirl and I'm impressed with what it can do in my terminal.

But again... I'm not thrilled to give it terminal access as the same user that I use to operate my machine on the daily. It being able to slurp ~/.ssh and such is not thrilling to me.

So, my first inclination was simply to install Claude Code under a different user, and run it as that user, so that it had access only to a home directory which was used for that purpose and no other.

...but then, one runs into the problem of editor integration. For example, I want to run Claude Code inside of VSCode or PyCharm, but these don't seem to readily run when launched from a terminal when using su as another user.

xhost gets VSCode to launch, but nothing works; it crashes as soon as one attempts to open a file.

So what are the current best practices?