Skip to content

SolutionsAI agents

Every agent’s computer, 24/7

A microVM of its own for each agent. It sleeps when the agent stops, and wakes in 3.5 ms with everything where it was.

Read the Sandbox docs

Python
from glac import Sandbox

sbx = Sandbox.create()                     # its own microVM
res = sbx.run("python3 -c 'print(6*7)'")   # res.stdout
sbx.files.write("/tmp/plan.md", "step 1")  # files in and out
snap = sbx.snapshot(name="step-1")         # memory and all
copies = sbx.fork(3)                       # 3 copies from now
sbx.sleep()                                # next call wakes it

Paid for while it waits

While the model answers, the agent’s computer has nothing to do. A sandbox sleeps after 30 seconds idle, and asleep it uses no CPU.

a month for the memory of a 4 GiB sandbox kept awake, at the list prices1
$23.36
for the same sandbox awake an hour a day and asleep the rest, keeping 1 GB1
$0.98

The life of a sandbox

Running while the agent works; asleep after 30 seconds idle, with everything in it; woken by the next command, file or request; saved whole while it runs, and forked into as many as 64 copies.

A sandbox called web: a computer of its own, worked from outside; a snapshot taken while it runs, and three forks going on from that moment.

What an agent gets

  • Kept apart: a microVM each, network rules, and secrets put in on the way out.

  • Where it left off: asleep, it keeps its memory, processes, disk and network place.

  • Work that lasts: a volume outlives the sandbox and goes wherever the agent works next.

  • People can step in: a terminal, a link to any port, or SSH with your own key.

  • Agents ready to run: Claude Code, Codex or OpenCode; the model’s keys are yours.

  • Every way in: the API, the glac CLI, five SDKs and an MCP server.

Read the docs

Illustrative previews

Measured, not promised

Medians on bare metal with nothing else running: a best case, and we say so. How we measure

to wake a sandbox that fell asleep by itself2
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 sandboxes once the service opens: an answer that had to wake one says how long that took, in its Server-Timing header.

Shell
curl -si https://api.glac.ai/v1/sandboxes/web/run \
  -H "Authorization: Bearer $GLAC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"command": "true"}' | grep -i server-timing
# server-timing: wake;dur=<milliseconds>

The products it uses

A month, worked out

A hundred agents for a month, at the list prices. Estimate your own

The use

  • 100 agents, each with a sandbox of 2 vCPUs and 4 GiB.
  • Each awake an hour a day for 30 days: 3,000 hours in all.
  • While awake, its CPU busy a tenth of the time, the rest waiting on the model; 2 GiB of its memory in use.
  • Asleep the rest of the month, each keeping about 1 GB: its memory image and its own layer of disk.
  • 10 GB sent to the internet.

A month at the list prices

$71.64

  • CPU2.16 billion CPU-ms × $0.01 / M$21.60
  • Memory6,000 GiB-h × $0.008$48.00
  • Traffic out10 GB × $0.004$0.04
  • Kept asleep100 GB-months × $0.02$2.00

On Paid this month is $73.60: the $5 fee, and $68.60 past what its inclusions cover, from your balance at the same prices.

Give your agent a computer

A key, then a sandbox from the command line or from Python, in a few minutes.

The MCP serverSandbox