A cloud server with Claude Code and Codex installed: your options

Four questions before the price

• Does it stay on? A coding agent on a long task prints nothing for minutes at a time, and several products decide a machine is idle by watching for traffic or terminal output. Read the idle rule on the product’s own page, and what happens to a running program when it applies.
• Whose sign-in do the agents use? There are three arrangements. You sign in on the machine with the maker’s own client, as on a laptop. You sign in once and the host keeps that login and hands it to each machine. Or the agent is pointed at the host’s model service and billed by the host. They differ in who holds your credential and whose bill the usage lands on. Your Claude or ChatGPT plan on a server covers what the labs’ terms say about a sign-in that someone else holds.
• Who can reach the machine? A public address with SSH open, a proxy in front, or a private network. And, separately, what the seller says its own side can do on it, which few pages state. Can your cloud host access your server covers that second half.
• What does the price cover? Almost always the machine and not the model: your Claude or ChatGPT plan is a separate bill. Then check whether it is a flat month or metered hours, a first-term price or the renewal, a virtual machine or a container, and whether you are root.

Everything below is from each company’s own pages on October 11, 2026. The products are examples of each shape, in no order of merit.

The labs’ own cloud: nothing to rent

If you pay for Claude or ChatGPT, you may already have this. Claude Code on the web runs each session in “a fresh virtual machine (VM) running Ubuntu 24.04” on Anthropic’s side, on Pro, Max and Team plans, with “no separate charge for the cloud VM”. Codex on the web does the same for ChatGPT Plus and above, each task in “its own isolated workspace”. Both start from a GitHub repository and end in a branch or a pull request.

What they are not is a server. Anthropic’s docs: “You don’t get a shell into the session VM. Claude runs every command for you”. A session’s machine pauses after a few minutes without activity, and “a paused VM can later be reclaimed”, taking any background work with it. OpenAI keeps a task’s saved state “recoverable for up to seven days”. For work that lives in one repository and finishes in a pull request, this is the answer and there is nothing to rent. The two guides on Claude Code’s cloud and Codex cloud set each against a machine of your own.

A plain server, or one with an agent on its image

The baseline is a server with nothing on it. Hetzner’s price list of June 15, 2026 has a 4 vCPU, 8 GB cloud server at €8.49 a month before tax in its cost-optimized line and €35.49 in its regular line; on October 11, 2026 its site marked the cost-optimized line as not available. DigitalOcean’s is $48 a month. Installing the two agents takes a few minutes each: Claude Code on Linux and Codex CLI on a Linux server have the commands and the errors. What you take on with the server is the rest: its firewall, its updates and its backups.

Some hosts save you the install with an image:

• RDP.sh sells a server at €5 a month (2 vCPU, 2 GB, 20 GB) whose image is “Ubuntu 24.04 with Claude Code and Codex CLIs, Node.js, Python with uv, Docker, tmux, headless Chromium and mosh”. You are root, it has a public address, and it offers a link that opens your terminal on any device with “No login needed” while you leave it switched on.
• Hostinger has a template named “Ubuntu 24.04 with Claude Code” for its ordinary servers. It carries Claude Code only; on its Codex page the buyer installs Codex. The 2 vCPU, 8 GB plan is $8.99 a month paid upfront and renews at $14.99.
• DigitalOcean’s Marketplace has a Codex image that is “pre-configured to use DigitalOcean Serverless Inference”, a model service DigitalOcean bills, with “Optional ChatGPT subscription auth via codex login”.

These are servers like any other: they stay on, you administer them, and they face the internet unless you close them. The image saves ten minutes on the first day. The DigitalOcean one is also a reminder to check the second question, since an image can arrive wired to a bill you did not expect.

Machines built for coding agents

This is the group the question usually has in mind, and it is where the four questions get different answers from product to product.

How you get a shell differs as well. exe.dev and boxd give you root over SSH, exe.dev through its own proxy with no public address on the machine. Sprites has its own command-line client and no SSH out of the box. CloudCLI lists SSH beside its browser and phone views, and boxes.dev offers SSH beside its apps. Cosyra is a phone app and a web client. Agent 37 gives a browser terminal and an API at a public address, and SSH with your own key. HostStack gives a browser terminal only: “There is no sshd in the image”.

Three patterns stand out once the rows sit side by side.

Half of them sleep, and the rule matters more than the price. Sprites says so outright: a Sprite “is not a server that runs until you turn it off”, and for its idle timer “redirecting to a file or detaching tmux doesn’t count” as activity. boxd warns about its own default: “The idle timers watch network traffic, not CPU”, so “A session that works with no traffic looks idle and the machine hibernates mid-job.” boxes.dev sleeps a machine after about five minutes without activity and says what holds that off: “An agent still working, you interacting with the machine, or visitors actively using a public link you’ve shared can keep it awake longer.” Sleeping when idle is right for many short tasks you start and watch, and it is why those products are cheap to leave around. For an agent you start at night and expect to find finished, read how the product decides a machine is idle, and change the setting where there is one. GitHub Codespaces, the general cloud development environment, has the same shape with no agent installed: it stops “after 30 minutes of inactivity”, and “A codespace has a maximum lifetime of 12 hours”.

Your sign-in is not always on the machine. On boxd you “Sign in once per organization, on any machine” and it serves that login to the rest. boxes.dev and HostStack keep the login on their side in the same way. exe.dev attaches its own model gateway to a new account as one way to supply the model, and a ChatGPT plan used through that gateway “is connected to exe.dev, not to the VM”. On CloudCLI, Cosyra and Agent 37 you sign in to each agent yourself. None of these arrangements is hidden: each page says which it uses. It is a thing to choose knowingly.

Few say what their own side can do. Agent 37’s privacy page does: “While we do not access your instance data in the normal course of operations, we do maintain infrastructure-level access to the servers that host your instance for the purposes of provisioning, updates, security, and support.” For most of the others, the product pages we read do not address it. That is a question to put to a seller before your code goes on its machine.

Two more differences hide under the word machine. Several of these are containers on a shared server, not virtual machines of your own, which limits what you can install and run inside. And “installed” does not mean kept current: exe.dev notes that its agents are installed when the image is built, so a long-lived machine keeps the version it was created with until you update it.

Everpod’s developer pod, by the same four questions

Everpod, which publishes this guide, sells a machine of that last kind. Its developer pod is a cloud computer of your own with your pick of Claude Code, Codex, OpenCode, Pi, Hermes and OpenClaw installed, from $24 a month for 4 vCPU, 8 GB of memory, 80 GB of disk. The four questions, answered in its own pages’ words:

• Does it stay on? A developer pod is one machine that stays yours: what you installed, the servers you left running, Docker and your work in progress are where you left them tomorrow.
• Whose sign-in? You sign in to each agent with your own subscription or API key. No model usage is included, and your sign-ins never pass through us.
• Who can reach it? It accepts no incoming connections from the open internet: no open ports and no public SSH. A server your agent starts is reachable from your devices and from nowhere else. By default the machine does not start connections to your other devices. Once you have joined the machine to your Tailscale network, switched off its key expiry and logged in, we remove the access we used to set it up. From then on Everpod has no login on it and no key. We do not log in to fix or update it: security updates install themselves, and the machine is yours to administer. We host the server it runs on, so we can restart it or restore one of its daily backups when you ask us to.
• What does the price cover? The machine, with you as its administrator. The whole disk is backed up daily, and the backups rotate over roughly 7 days.

You need a Tailscale account, which is Tailscale’s own and separate from Everpod. The machine joins your network with it, and the laptop or phone you log in from needs Tailscale too, signed in to the same account. If your laptop runs a consumer VPN, it may need to be off or set to let Tailscale through. What it asks of you is a terminal: you log in over SSH and run the agent there.

Which shape fits which job

• Your work lives in a GitHub repository and ends in a pull request. Use the cloud your Claude or ChatGPT plan already includes.
• You want the lowest bill and are at home on a Linux server. Rent a plain one and install the agents. An image saves the first ten minutes.
• You start many short tasks and watch them. A machine that sleeps when idle and bills by use is built for that.
• You want an agent to keep working for hours after you close the laptop. Choose something whose own page says it stays on, or a server, and check the idle rule before the price.
• You want one machine that stays on, with the agents on it and no public address, without building it. Look for those three things on the seller’s own page. Of the machines above, exe.dev’s pages describe them, and a developer pod is built to be that.