An open core

Agent Box is the runtime underneath.

Agent Box builds a working home for coding agents out of one config file: the agents installed and waiting for you to sign in, your files and sessions still there when you come back, remote control from Anthropic’s and OpenAI’s apps, and an address the outside world can reach. It’s the part of Station that does the work.

The source is on GitHub and the setup guide is in its documentation. MIT license, so there’s nothing to ask us for.

Station is the hosted version.

Station itself isn’t open source. It’s everything we run around Agent Box: the machine it runs on, the setup, the billing and somebody to email. What you’re paying for is not having to keep any of it running.

Your agent accounts stay yours.

Signing in happens inside your Station, through each provider’s own flow. Claude Code or Codex saves your login on your Station, the same way it would on your laptop, and Defang never receives it. The part of the runtime that holds it is the open part, so you can check that for yourself.

What we hold, and the one case that’s different →

You can take the runtime with you.

Agent Box installs from a one-click template on AWS or Azure, from a prebuilt image, or onto a Linux machine you already have. Whether that’s a better trade than paying us is a fair question, and we answer it on its own page.