What is an agent computer? Sandbox vs Computer vs VPS

An agent computer is a persistent, isolated machine that an AI agent owns. It keeps its files, installed tools, logins and services between sessions, and it is kept apart from your laptop, your production systems and other agents. A code sandbox is usually a short-lived box for running one piece of code, and a VPS is a server you run and maintain yourself. A Sanbox Computer is an agent computer: a persistent Firecracker microVM that you manage as a fleet from a CLI, an API and a web console.

Why does an agent need its own computer?

Coding agents and assistants don't just call a model. They run shell commands, install packages, clone repositories, start dev servers, log in to websites and keep notes. That work needs a machine, and the question is whose:

Your laptop holds your files, SSH keys, browser sessions and cloud credentials. An agent there can reach all of them, and it stops when you close the lid.
A production server holds real data and serves real users. An agent's mistake there becomes a production incident.
A machine of its own holds only what you put on it. It can keep working for days, and if something goes wrong you delete it and start again.

How is an agent computer different from a sandbox?

Sandboxes grew out of code execution: run a snippet the model wrote, return the output, throw the box away. They are built to start fast and live briefly, which is the right design for that job. Many sandbox products now also let you pause or stop a sandbox and keep its files, so it's worth checking what each one keeps and for how long.

An agent computer starts from the other end. By default it stays: the repository, node_modules, the browser profile and the agent's memory are still there next week. Short-lived copies are the special case, made from a prepared snapshot when you need many identical starting points.

How is an agent computer different from a VPS?

A VPS is persistent too, and you can run any agent on one. The difference is the work around it:

You install and patch the operating system, harden SSH, set up a firewall and arrange backups.
Outbound traffic is usually open, and the server usually has a public IP address.
API keys tend to sit in .env files or environment variables on the server, where the agent can read them.
One server often ends up hosting several agents and other services, so one agent's mistake reaches all of them.

A VPS is a good choice when you want full control and don't mind running a server. It gets harder when you need ten machines: one per task, one per teammate or one per customer.

Side by side

	Code sandbox (typical)	VPS	Sanbox Computer
Lifetime	Created per task; seconds to hours	Months, until you delete it	Until you delete it; stop when idle, resume later
What it keeps	Often nothing; some products keep files when paused	Everything on its disk	Its disk: `/workspace`, installed packages, Docker images, browser profile
Isolation	Container or microVM, depending on the provider	One VM, often shared by several services	One Firecracker microVM per Computer
Who maintains it	The provider	You	Sanbox runs the platform; you choose what to install
Network by default	Varies by provider	Open outbound, public IP	Outbound denied except what you allow; no inbound ports
API keys	Usually environment variables	Files or environment variables on the server	Secret Proxies keep keys outside the machine
Many at once	Through the provider's API	One by one, or your own tooling	Templates, snapshots, warm pools and forks from the CLI or API

What does a Sanbox Computer include?

Isolation. Each Computer is a Firecracker microVM with its own kernel and its own private disk, separated from other Computers by hardware virtualisation.
Persistence. Stop ends compute and saves the disk, and Resume boots the Computer fresh with that disk. Files, installed packages and Docker images survive. RAM and running processes don't.
Types. An OpenCode Computer runs OpenCode as a private server. A Browser Computer adds a managed browser you can watch and take over. A Hermes Computer is an assistant with memory, email and Telegram. A Docker Computer runs Docker Engine and Compose with no built-in agent. You can also install Claude Code or Codex in a Computer yourself.
Network policy. Outbound traffic is denied by default, and each template lists the hosts its Computers may reach. Inbound ports stay closed unless you add them, and nothing becomes public.
Secret Proxies. Sanbox adds a stored credential to matching HTTPS requests on their way out of the Computer, so the agent can use an API without seeing the key.
Fleet controls. A template defines a Computer once. Snapshots, warm pools and forks give you copies of a prepared state, and a Computer from a warm pool can start in under a second.
Data location. Computers, disks and snapshots are stored and processed in EU data centres.

Sanbox is sold to businesses and priced per organisation, with your own models and tools. Our forward-deployed engineers can build your first use case with you. Talk to us.

When Sanbox is not the right fit

You run thousands of short snippets and keep nothing. A code-execution sandbox or a plain container is simpler.
You need GPUs or Windows. Sanbox Computers offer neither.
You need macOS-only tools, such as iMessage or Xcode. A Mac is the right machine.
You want a server you fully control and you're happy to maintain it. A VPS does that well.

Sources

Firecracker, the open-source virtualisation technology that Sanbox Computers run on (checked 2026-11-06)