Agent sandboxes that sleep on the job—for free

What it does Dormice is a self-hosted sandbox platform for AI agents that runs on a single Linux machine as one daemon backed by a SQLite ledger. It treats sandboxes as permanent residents rather than disposable cloud resources, cooling them down through an automatic lifecycle that makes idle sandboxes nearly free to keep. An idempotent acquireSandbox call always returns the same sandbox, whether that means creating it, thawing it from a frozen state, or restoring it from S3.

The interesting bit The project inverts sandbox economics: measured on real hardware, freezing shrinks an idle sandbox holding 1 GiB down to roughly 5 MiB of resident memory, and waking it takes about 50 ms. It achieves this by suspending Docker and gVisor containers as processes, trading hardware virtualization for hibernation. The entire system is willfully constrained to one machine and one daemon—no Kubernetes, no external database, no fleet.

Key highlights

• e2b compatibility: the official SDK works by changing two URLs, making migration a configuration tweak rather than a rewrite.
• Honest lifecycle states: sandboxes move from active to frozen to stopped to archived, and the API explicitly reports when a restore is in progress rather than silently hanging.
• Single-binary deployment: one daemon, one SQLite database, one port. The web console, native SDK, CLI, and S3 archiver are all bundled together.
• Security defaults: the daemon binds to localhost only, network hardening for cloud metadata services is mandatory, and untrusted code runs under gVisor.
• Early but end-to-end: the full lifecycle from creation through freezing, stopping, archiving, and restoring, plus command execution, file I/O, and black-box e2b compatibility tests, all pass against real infrastructure.

Caveats

• Explicitly early development: the authors stress that nothing is ready for production yet, despite the full lifecycle working end to end.
• Linux-only and single-node by design: you need root access, Docker, gVisor, and a host tuned for high swappiness; multi-machine sharding is not implemented.
• No template builds: unlike the hosted e2b product, you bring your own Docker images and register them manually.

Verdict Worth a look if you run persistent AI agents and want to stop paying cloud sandbox rent for idle time. Skip it if you need a managed fleet, hardware virtualization, or a production-ready system today.