Sandboxes
A sandbox is a Firecracker microVM for one run of your agent’s code. It boots its own kernel, starts with the network off, and goes away when you stop it.
What a sandbox is
Each sandbox is a separate microVM with its own kernel, not a container that shares a kernel with other work. It starts from a template, runs the commands you send it, and keeps its files until you stop it.
| You can | How |
|---|---|
| Start one | sandscale sandbox start --template TEMPLATE |
| Run a command and read its output | sandscale sandbox exec ID -- COMMAND |
| Move files in and out | sandscale sandbox files put, files get, files ls |
| Open a shell | sandscale sandbox connect ID |
| Pause it and resume it later, with memory and files kept | In the console or the API |
| Save a snapshot to start new sandboxes from | Pro and up |
| Stop it | sandscale sandbox stop ID |
Templates
A template is the image a sandbox starts from. sandscale sandbox start --options lists the templates, sizes and time limits your plan and your organization’s policy let you use. You can build your own with sandscale sandbox template init and sandscale sandbox template build.
Network
A new sandbox has no network. On Pro and up you can ask for network access when you start it (--network), and it can reach only what your organization’s egress allowlist allows. Free sandboxes always run with the network off.
Sizes and limits
How many sandboxes you can run at once and how many vCPUs each can have depend on your plan:
- Free: 2 at once, up to 2 vCPU each.
- Pro: 20 at once, up to 8 vCPU each.
- Team: 100 at once, up to 16 vCPU each.
- Enterprise: up to 64 vCPU per box on request.
Running time counts against your plan’s vCPU-hours. See Plans and limits.
Pause, resume and snapshots
Pausing keeps a sandbox’s memory and files, and resuming brings it back where it stopped. Every plan can pause and resume. Snapshots save a sandbox so you can start new ones from that point; they use storage, so they come with Pro and up.
From your agent
With sandscale mcp, your coding agent can list, start, run commands in and stop sandboxes itself, with your sign-in and inside your policy.