Host Codex

Why this one

Pick this when your team is already standardised on OpenAI models and you want the coding agent to match. Same shape as hosted Claude Code — persistent workspace, Telegram front door — with OpenAI's harness behind it.

All 9 hosted coding agents share the same shape — a full agent with files, a shell and git, running in its own Firecracker micro-VM with a persistent workspace, reachable from Telegram and the other channels listed above. If you have not settled on one yet, the side-by-side comparison is the faster read.

What you need before you start

A key. An OpenAI API key, a ChatGPT plan linked via the subscription flow, or the included managed model.

That is genuinely the whole list. There is no server to provision, no Docker to install, and no reverse proxy to configure — the point of hosting it here is that the micro-VM, the volume and the channel bridge already exist.

Setting it up

1. Describe the agent. Say what you want it to do in plain language. Naming Codex explicitly is fine, and the button at the end of this page does it for you.
2. Add your key. The builder asks for exactly the credential above and seals it — secrets are AES-GCM encrypted into the machine config and never stored in plaintext by the control plane.
3. Deploy. One click provisions the micro-VM. First boot is the slow one; after that the agent stops when idle and wakes on the next message.
4. Connect a channel. Telegram is the usual front door — you paste a bot token from BotFather and the bridge does the rest. Codex also speaks slack, discord, email, matrix, whatsapp, sms.
5. Talk to it. Web chat works immediately and is the fastest way to check the key and the model are right before you wire up a channel.

The thing that will bite you

Codex is deliberately conservative about running commands it was not asked to run. That is usually what you want on a box that holds a git workspace, but it means fewer things happen without an explicit instruction than you may expect.

What it costs

Two separate bills, and only one of them is predictable. Hosting is $9 a month for the Standard box (1 vCPU / 1 GB), fixed, and the first agent month is free. Model tokens are forwarded at provider cost with no markup, so that half depends entirely on how much you use it.

The variable half is where always-on agents surprise people, and it is worth reading the cost breakdown before you point a coding agent at a long-running task. A coding agent that gets stuck in a retry loop is the single most expensive failure mode on this platform, which is why cost per conversation is on the Billing tab rather than buried in an invoice at the end of the month.

Getting the work out

The workspace lives on a persistent volume, so files survive restarts and idle stops. If you want the agent's output somewhere other than the VM, give it a git remote it can push to — that is the difference between an agent whose work you can use and one whose work you have to copy out of a chat window by hand.

Is Codex (hosted) the right pick?

Not sure? Describe the job rather than the runtime — the builder's solutions engineer reads the same capability matrix this page does and will argue for a different one if it fits better.