An agent needs a computer, not a function
A model that writes files, installs packages and runs commands needs a machine. Sandboxes give it one: an isolated cloud computer that starts in about a second, bills by the second, and goes away when the work is done.
One task cannot reach another, and an untrusted script cannot reach you. That is the whole promise, and it is what a serverless function could never give an agent.
One run started 32,752 of them
To find the ceiling we painted with it. Each pixel of the Munito wordmark and its cube became one sandbox: 32,752 machines running at once, started through the public API from Paris, on our own European capacity. Watch the run.
The same call starts the first one for your agent.
What you get
npm install -g @munitoai/cli
munito create --size small --timeout 600A sandbox has a shell, a browser, SSH, open ports and a desktop, so an agent works the way a person would. Reach it from the command line tool, from TypeScript, or straight from the API.
Pick the template that fits the job: one for plain code execution and CI, one that carries the coding agents, one with a desktop. See Templates.
Keep the disk between jobs
Build a sandbox with your toolchain, pause it, and fork it for each run. A fork starts with the dependencies already installed, which is what a CI job and a reinforcement-learning rollout both need, and it is where the minutes go when every run rebuilds the same image.
Where it runs
In the EU, on the same account and the same key as the inference API, with per-second rates and no monthly floor. Sandboxes is a public preview, so tell us what breaks.
See Sandboxes, the Sandboxes guide and what it costs, or talk to an engineer.