AI coding agents like Claude Code and Codex pose security risks when run with unrestricted access on developer machines. MicroVMs offer a practical middle ground between containers (which share the host kernel) and full VMs. Using Fedora Linux, podman, and the krun runtime, developers can run each agent in its own microVM with its own kernel, providing stronger isolation. The article walks through a complete setup including a Dockerfile, docker-compose.yaml with krun annotations for UID/GID translation and SELinux labeling, and an entrypoint script. Key gotchas include allocating sufficient RAM/CPU, using libkrun >= 1.8, and handling the USER directive manually. A helper script automating the entire setup is also provided. The author notes microVMs are not bulletproof and recommends full VMs for truly dangerous workloads.
