I run EVERYTHING inside MicroVMs with LimaVM. Employer pays for Claude Code, which I run inside lima with herdr (because I often work on 5-7 different worktrees). I also pay for my own Claude Code subscription, and run each project in a dedicated limavm with plain `claude --dangerously-skip-permissions`. All this inside `tmux`. For secrets protection I run `agent-vault`[1] so claude does not have access to API keys/secrets. I try not to run any npm/cargo/ruby stuff on my machine, but run everything from the vm. If (when) the VM will get compromised, I just wipe it off, and restart from scratch.
I use MCPs here and there (google search console, sentry). Claude helped me to configure `tmux` to show alert popups when one of the clauds finished its work or needs my input. I utilize worktrees as much as possible to work on stuff in parallel. I have a fancy neovim setup that I no longer use, unfortunately. When I feel like reviewing the code, I use `tuicr`[2]. So basically, I live inside my terminal and firefox.