Welcome - boxd

If you’re an AI agent reading this, fetch llms.txt or llms-full.txt. They’re built for you.

Persistent cloud compute for the agent era.

Agents run for hours. They fork their paths. They need to remember. Today they get a choice between two bad shapes. Containers and Kubernetes pods are built for one operator. Slow boot. Always-on bill. No fork. No isolation between agents. Ephemeral sandboxes go the other way. Built for one call. No memory. Dies in seconds. The agent loses its work between turns. So every serious team shipping agents builds the missing harness themselves, and pays for it twice. Once to build. Then forever, to maintain. boxd ships persistent computers as a primitive. Three operations.

fork

Branch the entire state of a running computer. Memory, disk, in-progress work. ~160ms.

suspend

Freeze when idle. Disk and memory persist. No idle bill.

resume

Wake any computer, anywhere on the fleet. Sub-millisecond.

Built on a stack we own from KVM up. ~50ms cold boot. 4 to 100× faster than anything stitched from Firecracker, Kubernetes, and Postgres. Single binary. Zero hyperscaler dependencies. EU-sovereign infrastructure. Same machine model for your dev box, your PR preview, and the sandbox you hand to an agent.