Using AI commandline tools can require allowing some scary permissions (ex: "allow model to rm -rf?"), I wanted to isolate commands using a VM that could be ephemeral (erased each time), or persistent, as needed. So instead of the AI trying to "reason out" math, it can write a little program and run it to get the answer directly. This VASTLY increases good output. This was also an experiment to use claude to create what I needed, and I'm very happy with the result.
Show HN: Created a script to run commands in an isolated VM for AI tool calling | Better HN
