Skip to content

SolutionsVibe coding

Every prompt, a running app

Your agent builds each app in a sandbox, with a link to preview it, then ships it as a Cube that sleeps until someone opens it.

Read the Cube docs

Python
from glac import Glac, AGENTS

glac = Glac()
sbx = glac.sandboxes.create(image=AGENTS, env=model_keys)
sbx.agent("claude").run("Make a todo app on port 3000")
print(sbx.url(3000))                  # preview: a link to the port
image = "ghcr.io/you/todo:1"          # what your build pushed
app = glac.cube.deploy(image, name="todo", port=3000)
print(app.url)                        # its own HTTPS address
app.rollback(1)                       # back a version in one call

Two machines for every app

Every app needs a machine to build in and one to run on, each paid for all month if kept running. Here both sleep: the sandbox when the agent stops, the Cube between requests.

a month for the memory alone of 1,000 apps of 512 MiB kept running, at the list prices1
$2,920.00
for the CPU and memory of an app that nobody opens: it is asleep1
$0

Build in a sandbox, ship as a Cube

The agent works in a sandbox, with a link to its dev server and a fork for every attempt. What it ships is an image, deployed as a Cube at an address of its own, with instances that wake for requests.

Build: the agent’s sandbox, a snapshot of it, and forks that try again from that moment.
Ship: the app’s versions, the ready one’s template, and its instances in each region, awake for a request and asleep after it.

What your product gets

  • Build in a sandbox: a terminal, a link to the dev server, a fork for every try.

  • A volume for the database, billed by the data it holds.

  • Ship as a Cube: every deploy a version, at its address or your customer’s domain.

  • Asleep until opened, 100 ms after its last request, and awake with its memory.

  • One instance per end user, picked by a key in the request.

  • Usage per customer, read by their label, with a monthly spend limit.

Read the docs

Illustrative previews

Measured, not promised

Medians on bare metal with nothing else running: a best case, and we say so. A Cube’s instance is the same kind as a sandbox. How we measure

to wake a sleeping instance2
3.5 ms
for the call that woke it, sent to answered2
6.7 ms
to put it to sleep, with its memory and processes2
2.3 ms

See it for your own apps once the service opens: the log’s line for a request that woke an instance says how long the wake took.

Shell
glac cube logs todo -f
# one line a request; one that woke an instance says how long it took

The products it uses

A month, worked out

A thousand apps made and run in a month, at the list prices. Estimate your own

The use

  • 1,000 apps, each built by the agent in a sandbox of 2 vCPUs and 4 GiB for half an hour, its CPU busy a fifth of the time, 2 GiB in use; the sandbox deleted after.
  • Each app then a Cube answering 30,000 requests in the month, 20 ms of CPU each, 256 MiB in use, 20 KB an answer.
  • Requests arriving at random; an instance awake while it answers and for 100 ms after.

A month at the list prices

$34.60

  • CPU1.32 billion CPU-ms × $0.01 / M$13.20
  • Memory1,250 GiB-h × $0.008$10.00
  • Requests30 million × $0.30 / M$9.00
  • Traffic out600 GB × $0.004$2.40

On Paid this month is $31.60: the $5 fee, and $26.60 past what its inclusions cover, from your balance at the same prices.

Ship the first app

A key, then an image deployed as a Cube with one command, or a sandbox for your agent to build in.

CubeSandbox