Post

Experimenting with Firecracker microVMs at work for developer environments. So far I have: - Developers login with GitHub and get a box - Box is reachable via SSH and will auto-start - Box can be stopped/recreated from UI Each box - 32G ram - Mise with tools/agents pre-installed - Paseo relay

Andri Óskarsson

@andri.dk

Host machine and architecture - Nobody needs to access it - VM API is a non-root user w. KVM access - Management CLI for setting up systemd/firewall - Services run via SystemD - Web interface is separate from VM management Stack is Go and a little SQLite

Andri Óskarsson

@andri.dk

Developers can then use: - SSH client and run Claude/Codex/etc from there - Zed via SSH - Visual Studio Code Remote - Jetbrains IDE w. remote SSH or whatever they want, as long as it can connect via SSH

14/10/26, 6:28 am0 reposts9 likes

Andri Óskarsson

@andri.dk

Non-goals: - Multiple host-machines or a cluster scheduler. I'm not building Kubernetes! - Anything that unnecessarily increases the exposure of the VM manager API

Andri Óskarsson

@andri.dk

Maybe later: - Multiple VMs per developer, identified by SSH username - Killing idle machines automatically - More UI controls - Transparent HTTP proxy w. logging and auto injected CA certificates - Secret management

Chraz

@edu.runlinea.com

Monday, one click: recreate the box. Yesterday's agent that cornered itself with sudo is gone with the microVM. Copyable bit: the VM API stays non-root, so Claude never sees the host.