Skip to content

Sandbox

A machine of its own for each AI agent, user or job, made from any OCI image in milliseconds. Run commands, read and write files, open a terminal; it sleeps when idle and wakes with its processes running.

Updated View as Markdown

A sandbox is a virtual machine of its own: a Firecracker microVM that runs an OCI image, the kind docker builds and registries serve. You make one with an API call. Then you run commands in it, read and write its files, open a terminal or a link to one of its ports. Commands and files reach it through an agent inside the machine, not through its network.

When nothing happens in it, a sandbox sleeps. It keeps its memory, its processes and its disk, and the next call wakes it where it was, in milliseconds. It stays until you delete it.

When to use Sandbox

  • An AI agent’s computer: a coding agent, a browser agent or a code interpreter that runs commands, installs packages and keeps its work between turns.
  • Code you did not write: what users or models send you, kept away from your own machines and from each other, with network rules and secrets the code never sees.
  • A job per request: glac run makes a sandbox, runs one command and deletes it.
  • Environments that branch: a snapshot of a running sandbox, and copies (forks) that each go on from that moment.
  • A server of its own: with init: "systemd" the image’s own systemd is its PID 1, for packages, services, cron and SSH as on any server (a whole system).

To serve an image at an address, with instances that start for requests, use Cube. Both run on the same machines, at the same prices.

How it works

  1. Make. POST /v1/sandboxes (glac sandboxes create) with an image, or with none for the platform’s default image. The first sandbox of an image waits while the platform pulls it and saves a template. Later ones are copies of that template, from a pool that is already running.
  2. Use. Run commands (exec), read and write files, open a terminal, give out links to ports. Calls wait for a sandbox that is starting or asleep: none fails for that.
  3. Sleep. After sleep_after_seconds idle (30 unless you say), it sleeps. Whatever needs it wakes it: a command, a file, a request on a link, a timer.

What it costs

A sandbox pays for the CPU it uses and the memory it holds while it is awake. Asleep, it pays for what it keeps: its memory and the part of its disk that is its own. One held paused in the machine’s memory pays for that memory as if awake, and no CPU.

Meter Price Free, each day Paid includes a month
CPU, while it computes $0.01 a million CPU-ms 1M CPU-ms 60M CPU-ms
Memory, while awake or paused in memory $0.008 a GiB-hour 6 GiB-hours 250 GiB-hours
Requests, into Cubes and through sandbox links $0.30 a million 100,000 10 million
What a sleeping instance keeps (its memory and its own disk) $0.02 a GB-month 15 GB of data in all, with snapshots and Drives 20 GB-months of data, with snapshots and Drives
What snapshots keep $0.02 a GB-month The same data The same data
Traffic out to the internet, from North America $0.004 a GB 3 GB, from any region 1 TB, from any region
Traffic out to the internet, from Europe $0.004 a GB
Traffic out to the internet, from Asia-Pacific $0.01 a GB

A paid plan includes no quantity of a meter: its fee comes back as the same amount of credit, which pays for any meter at these prices. See pricing and plans.

Limits of each plan

Free Paid
Largest instance 2 vCPUs, 4 GiB 8 vCPUs, 16 GiB
Instances awake at once 5 500
Their vCPUs together 10 500
Their memory together 10 GiB 1,000 GiB
Sleeping instances kept 20 10,000
Sandboxes made a minute 10 600
Sandboxes one project makes an hour 200 10,000

The sandboxes awake, their vCPUs and their memory are counted for the whole account: sandboxes and Cubes’ instances of every project together. Every limit is on the limits page.

Next

Navigation

Type to search…

↑↓ navigate↵ selectEsc close