Skip to content

Next-generation computing for AI agents, web apps, APIs, sandboxes, remote development, CI/CD, databases and scraping

Cube runs any container image in a microVM of its own: asleep when idle, awake in 3.5 ms.

Talk to us
  • 3.5 ms to wake1
  • 2.3 ms to sleep1
  • CPU billed per busy millisecond
  • Any container image

Isolation

A computer for every program

Every app, agent and browser is a Firecracker microVM of its own, booted from any container image. Its own kernel and its own memory, and everything in it belongs to one project.

Sleep

Asleep and awake in milliseconds

to wake1
3.5ms
to sleep1
2.3ms
idle, then a Cube sleeps
100ms

Woken by a request, it picks up where it left off, with its memory, processes and connections. Stateless or stateful, it is your choice.

Global

Global by default

One deploy runs in every region we operate, and each request is answered in the region nearest to it.

Live
AshburnUnited StatesCoventryUnited Kingdom
Being built
New York · Frankfurt · Finland · Singapore · Japan

Cube, and the applications run on it

The platform

Cube

Any container image as an app that sleeps between requests

Managed Applications

Sandbox

A computer of its own for every agent

Soon

Managed Applications

Browser

Real browsers, driven from code

Soon

Managed Applications

Serverless Postgres

Postgres that sleeps when idle and wakes on a connection

Soon

Managed Applications

Third-party apps

Open-source projects run for you, then a marketplace

Versions

Every deploy is a version; roll back in one call.

HTTP/1.1, HTTP/2, HTTP/3 and gRPC

Volumes

Disks that outlive their instance, billed by their data and their cache.

Up to 1 TiB each, for one at a time or shared

Your own domain, its certificate issued and renewed for you.

An instance per key, with its memory and disk.

A private network for each project.

For everyone who ships

Built for people and their agents

An agent, through the MCP server
  1. Calls sandbox_run sandbox: "web", image: "python:3.12", command: "pip install -r requirements.txt && pytest"
  2. Answer: exit 0 · 42 passed in 3.1s
  3. Calls sandbox_snapshot sandbox: "web", name: "before-upgrade"
  4. Answer: snapshot before-upgrade, the sandbox running on
  5. Calls cube_deploy app: "api", image: "ghcr.io/acme/api:1.4"
  6. Answer: ready · https://api-‹project›.glac.run
  • Developers

    Any container image, your Compose file imported in one command, a command line that starts in 3 ms and SDKs in five languages.

  • Teams and companies

    Accounts with owners, admins and members, projects with keys of their own, an audit log, spend limits, a DPA and an SLA.

  • AI agents

    An MCP server with 73 tools: an agent makes its own sandboxes, runs its code, ships what it built and cleans up.

Pay for what runs

CPU is billed per busy millisecond, never while it sleeps. Asleep, you pay only for the state it keeps.

  • $0.01a million CPU-msBusy CPU, by the millisecond
  • $0.008a GiB-hour of memoryWhile in the host’s memory
  • $0.30a million requestsInto a Cube or a sandbox’s link
  • $0.02a GB-month kept2Volumes, snapshots, sleep
  • $0.004a GB of traffic outFrom North America and Europe

One price list for everything: Cube and every Managed Application are billed by these meters.

Free
$0, no card, quotas each day
Paid
$5 a month, inclusions each month

Every price and limitEstimate a month

Questions

The rest is in the docs.

What can I run?

Any OCI container image, public or private, as a Cube: a service at its own address that wakes for requests, or a job that runs to its end. A Volume keeps what outlives an instance. For agents, a Sandbox is an instance you drive through the API: commands, files, snapshots.

What happens when it sleeps?

Left idle, a Cube’s instance sleeps 100 ms after its last request and a sandbox after 30 seconds, unless you say otherwise. It keeps its memory, its processes, its disk and its place on the network, and the next request wakes it right where it left off.

Where does it run?

Today in two data centres: Ashburn and Coventry. Being built: New York, Frankfurt, Finland, Singapore and Japan. Every Cube runs in every region unless you choose fewer, and each request is answered in the region nearest to it. A sandbox runs in the one region you make it in.

How is my code kept apart from others?

Every instance of a Cube, and so every sandbox, is a Firecracker microVM of its own. Commands and files reach it through an agent inside the machine, not through its network, and everything in it belongs to one project.

What are Managed Applications?

Software we run for you on Cube: Sandbox today, and Browser, Serverless Postgres and open-source projects next. Each is billed by the Cube resources it uses, its CPU, memory, traffic and storage, at the same prices as everything else.

What do I pay while it sleeps?

No CPU time. Only the state it keeps: its saved memory and the part of its disk that is its own, by the GB-month; or, while it sleeps paused in the host’s memory, that memory at the price it has awake.

What do Free and Paid include?

Free needs no card and has quotas of each meter every day, back at 00:00 UTC; it is never charged, and a quota used up stops that one thing until then. Paid is $5 a month and includes a quantity of each meter every month; past it, each meter is paid at the list price from your balance.

Start building with Glac

Start on Free. No card needed to try it.

Talk to us