Firecracker micro-VMs for AI agents
Give every agent
a computer of its own.
Run untrusted or model-generated code inside a Firecracker micro-VM with its own kernel, governed egress and a lifecycle built for agent work.
A real machine boundary
Not a shared kernel you hope holds.
Each workload gets a Firecracker micro-VM with its own guest kernel. Code inside one sandbox cannot see the host or another tenant.
One tool call at a time
Spawn. Run. Release.
The lifecycle maps directly to how an agent works: provision a machine, execute the task, then destroy or preserve it.
- SPAWN01
createSandbox()Provision the boundary
Choose a compute shape and root filesystem. The workload receives its own micro-VM.
- RUN02
runCommand()Execute inside it
Drive the task from TypeScript, the CLI or MCP and read the result from the guest.
- RELEASE03
destroy() / pause()End compute cleanly
Tear it down when finished, or snapshot the working state for the agent’s next step.
State survives the idle gap
Pause the bill.
Keep the machine.
Pausing serializes guest memory and device state to disk, stops compute billing and restores the same working session into a fresh micro-VM when the agent returns.
Real architecture. Clear stage.
CreateOS Sandbox publicly identifies the product as alpha. Audit logging, RBAC and major compliance certifications are on the roadmap—not claimed as already held.
For teams that need sovereignty now, the documented interim control is self-hosting the control plane and storage inside their own boundary.
Your agent writes the code
Give it somewhere safe to run.
Start with free credits, provision one micro-VM and inspect the boundary from the first command.
Spin up a sandbox