This is really impressive work on the infrastructure side.
Skipping hyperscalers like AWS for Hetzner/OVH is definitely a massive lever, but your choice of gVisor for memory overcommit is where this gets technically interesting. A few questions on the implementation:
• Memory & Density: Since gVisor allows you to share memory without reserving a full VM's allocation, how are you handling noisy neighbors if multiple sandboxes spike in RAM usage simultaneously? Are you using specific cgroups configurations or out-of-memory (OOM) killer tuning to prevent one sandbox from crashing the host?
• Storage Persistence: You mentioned that files stay with the agent across sessions. How are you handling the storage layer efficiently at $0.05–$0.10/user? Are you using localized NVMe storage with networked backups, or something like ceph/block storage?
• Cold Starts: When a sandbox wakes up from auto-sleep, what is the exact latency for the user?
Congrats on the launch and the pivot from a simple wrapper to a real infrastructure play!