Why we built Station
We built Station because our agents needed a place that stays: the app keeps running, both agents share the work, and what they build ships to your own cloud.
The first Station was built for a job we already had. We wanted Defang Deploy to run more of its own product work through agents: read an issue, make the change, review it, and ship through the same path our customers use. That sounds simple until the agent doing the work is tied to one laptop. Close the lid and the work stops. Nothing can reach it, and the preview it just built goes with it.
Founders hit the same wall from the other side. They build something real in an app builder, outgrow the credits and the guesswork, and move to Claude Code or Codex on a real repo. The work gets better straight away. Then they find the agent is only as available as the laptop it runs on.
Why not use Anthropic’s or OpenAI’s cloud?
When we started building Agent Box, the runtime under Station, in July 2026, Claude Code’s cloud sessions were still a research preview, and Codex’s cloud ran each task in a fresh sandbox. We expected that when Anthropic and OpenAI shipped versions that keep going, neither would run the other’s agent. And we already wanted more than a place to run one: a way to put what it builds online at a link, to receive webhooks, and to ship to production.
Both have shipped since. As of late September 2026, Claude Code’s cloud sessions are generally available and Codex has reusable cloud environments. Both keep working when you close your laptop, and both are good for handing off a task and getting a pull request back. But each new session starts fresh from your saved setup, with nothing still running from the last one, and, as we expected, each cloud runs only its own agent. A product needs more than that: the app, its database and the preview you sent a customer still need to be there tomorrow.
One Station, both agents, and nothing resets between sessions.
Here’s what that means in practice. The side-by-side, with links to their docs, is on the comparison page.
It keeps running
A Station is a real computer that doesn’t go away. Your files, your tools, your signed-in agents and your running app are where you left them. Ask your agent to publish what it built and you get a link you can open on your phone or send to a customer. It stays up because your Station does.
Both agents, on your own accounts
Claude Code and Codex work side by side on the same Station and the same files. Each maker’s cloud runs its own agent.
Station doesn’t put its own model in front of you. You sign in to the official Claude Code or Codex inside your Station, with your own Anthropic or OpenAI account. Claude Code or Codex saves your login on your Station, the same way it would on your laptop, and Defang never receives it. We don’t proxy your agent traffic and we don’t resell tokens, so your AI bill stays between you and your provider. One exception you should know about: a value you type into your Station’s environment settings yourself does pass through Defang on its way there. The security guide spells out both.
From “it works” to production
Claude Code and Codex cloud tasks end in a pull request, and deploying it is up to you. Station is where your agents work. Defang Deployis how what they build reaches real infrastructure: dev, staging or production, on the cloud and region you choose, in your own account. A Station itself runs on Defang’s cloud. The thing that reaches your cloud is the deployment.
Once it’s live, you become the relay. Something breaks, you find out, you open your laptop, you paste the error to your agent. Ask your agent to connect your Station to GitHub, Linear or your own app, and the event reaches it directly, on the Station where the code already is. Their clouds can start work from events too, from GitHub, a schedule or an API call. The difference is where it lands: on the Station next to your running app, not in a fresh session that clones the repo. The agent gets told and can start work. It isn’t promised to fix production while you sleep, and you still review what it ships. You turn up for the review instead of the relay.
Built on an open runtime
The part of Station that handles your sign-ins, keeps your sessions alive and connects them to Anthropic’s and OpenAI’s apps is Agent Box, and it’s MIT licensed. You can read exactly how it handles the accounts you connect, and if Station ever stops earning your trust, you can take that runtime and run it yourself. To be precise: Agent Box, the runtime, is open source. Station, the hosted product around it, is not. More on what is open.
What it doesn’t do yet
One Station per account, today. If you’re managing a team of agents across a company, we’re not there yet. Starting one isn’t instant either: building a real machine, booting it, attaching storage and wiring its address usually takes about two minutes. You connect your accounts while it builds, so the wait does something.
Stations run on Defang’s cloud, on Azure, in the US, Canada or the EU, whichever serves where you sign up from. If you need somewhere else, self-hosting Agent Box is one answer. And Station doesn’t make the agent smarter than it already is. It still needs direction, and you still review what it ships. That’s how the product works right now, and we don’t plan to quietly drop any of it from the copy later.
What it costs
The first hour is free, no card. After that a card starts a 7-day trial, one per customer. If you’ve had one before, the $39 starts straight away. Otherwise nothing is charged until the trial ends. One plan: $39/month, flat, per Station. No credits, no token markup, no usage meter to watch.
A bare VPS is cheaper, and if you enjoy maintaining a server, a $10 one running the open-source Agent Box runtime is a reasonable setup: it’s the same code. $39/month buys the version that’s set up for you and runs on our cloud account, plus a path to actually shipping what gets built.
“The next version of HIO is a mobile app, a cloud back end and an ML recommendation engine ~ and with Defang's help, I'm the one holding all of it together. Station is where that work happens. I can see what's in flight, get features built and out to production, and when something breaks in production I can go find it myself instead of waiting on a team of devs.”
We built what we wanted once our agents needed more than a laptop, and more than a task that ends. If that’s where you are too, try Station.