Run OpenAI Codex on an always-on Linux computer

Install the Codex CLI on a Linux machine that stays on, sign in with a device code or an API key, and run it inside a terminal session that keeps going when you disconnect. OpenAI's desktop app can now open projects on an SSH host, but that host is still a machine you have to run. On Sanbox, you install Codex on an OpenCode Computer, which stays running until you stop it, keeps its disk across stops, and is reachable from the console's browser terminal or sanbox ssh.

Run this with your agent

Read https://sanbox.cloud/kb/coding-agents/codex-on-an-always-on-linux-computer.md and follow it to run Codex on an always-on Computer.
If the Sanbox CLI is missing, install it with `npm install -g @sanlabs/sanbox-cli@latest`
and sign in with `sanbox login`. Show me each command before you create, stop or delete a Computer.

What did people ask for?

In openai/codex#10450, "Remote Development in Codex Desktop App", users asked to work on code that lives on remote hosts, as VS Code Remote SSH allows. It drew 879 reactions and was closed as completed in May 2026. OpenAI's docs now describe remote connections: you add the host to ~/.ssh/config, install and sign in to Codex on that host, and the desktop app runs commands and edits files there. Phone control through the ChatGPT mobile app goes to a macOS or Windows machine running the desktop app, which has to stay awake.

So you still need an always-on Linux machine. The steps below use a Sanbox Computer; the Codex parts work the same on any server.

Steps

1. Allow the hosts Codex needs

Save this as codex-policy.json:

{
  "default_action": "deny",
  "rules": [
    { "destination": "registry.npmjs.org", "ports": [{ "from": 443, "to": 443 }] },
    { "destination": "api.openai.com", "ports": [{ "from": 443, "to": 443 }] },
    { "destination": "auth.openai.com", "ports": [{ "from": 443, "to": 443 }] },
    { "destination": "chatgpt.com", "ports": [{ "from": 443, "to": 443 }] },
    { "destination": "github.com", "ports": [{ "from": 443, "to": 443 }] }
  ]
}

registry.npmjs.org serves the install and api.openai.com the API. auth.openai.com (sign-in) and chatgpt.com (requests on your ChatGPT plan) appear in Codex's source; you need them only for ChatGPT sign-in. Check this list against OpenAI's current docs; a blocked request names the host it couldn't reach. github.com is for cloning.

2. Create and start an OpenCode Computer

The image has Node, npm, git and the zellij terminal multiplexer. Its own OpenCode agent needs a model provider from your organisation:

sanbox model-providers list --json
sanbox templates create --name "Codex" --type opencode-computer \
  --model-provider <provider-id> --model <model-id> \
  --network-policy codex-policy.json --json
sanbox templates validate <template-id> --json
sanbox run --template <template-id> --external-run-id codex-main --json
sanbox runs get <run-id> --json

Wait until the state is running. Don't use --wait on sanbox run; for a Computer it waits until the Computer stops.

3. Install Codex

sanbox exec <run-id> --root -- npm install -g @openai/codex
sanbox exec <run-id> -- codex --version

OpenAI also publishes a standalone installer for macOS and Linux (curl -fsSL https://chatgpt.com/codex/install.sh | sh). The npm package is used here because the policy already allows npm. Update Codex later by rerunning the same install command.

4. Open a terminal that survives disconnects

Use the console's browser terminal: each terminal is a zellij session that keeps running when you close the tab or lose the network, and you can reopen it from another device. Or, from your own terminal:

sanbox ssh <run-id>

Then, inside the Computer:

zellij attach --create codex

If your SSH connection drops, run sanbox ssh <run-id> and zellij attach codex to get back. The console terminal and SSH can use different home directories, which is where Codex keeps its login, so sign in from the one you'll keep using.

With your ChatGPT plan, use a device code. Turn on device code login in your ChatGPT security settings first (workspace admins allow it in workspace permissions), then:

codex login --device-auth

Open the link on any device, sign in and enter the one-time code. OpenAI marks device code sign-in as beta.

With an API key, which OpenAI recommends for automation, read it without echoing it and pipe it in:

read -rs OPENAI_API_KEY && export OPENAI_API_KEY
printenv OPENAI_API_KEY | codex login --with-api-key
unset OPENAI_API_KEY

Codex caches the login in ~/.codex/auth.json or your system's credential store. Treat that file like a password.

6. Run Codex

cd /workspace
git clone https://github.com/<owner>/<repo>.git
cd <repo>
codex

Close the tab or the SSH window; Codex keeps working in the zellij session.

7. Stop and resume

sanbox runs stop <run-id> --wait --json
sanbox runs resume <run-id> --wait --json

Stop ends Codex and every other process. The disk is kept: the repository, the Codex install and its login. After a resume, open a terminal and start Codex again.

Lock down egress

If you sign in with an API key, remove auth.openai.com and chatgpt.com. Add package registries your project needs, such as pypi.org, as separate rules.

Snapshot it

Stop the Computer after installing Codex and cloning the repository, but before you sign in, and save it. New Computers from the snapshot start ready, without your login:

sanbox runs stop <run-id> --wait --json
sanbox snapshots create --run <run-id> --name codex-base --wait
sanbox run --snapshot codex-base --json

What this doesn't do

Sanbox doesn't document using a Computer as an SSH remote host for OpenAI's desktop app. Run the Codex CLI inside the Computer instead.
The ChatGPT mobile app's remote control doesn't reach the Computer. Use the console terminal in your phone's browser, and check how it handles your phone's keyboard.
It doesn't keep processes through a stop, and nothing stops or schedules it for you. A running Computer stays on until you stop it.
Your Codex login is a file on the Computer's disk. Forks and snapshots copy it, and anyone in your organisation who can open the run's terminal can use it. Plan logins are personal: on a Computer several people share, use an API key.
Model usage is billed by OpenAI under your plan or API key. Your Sanbox plan, agreed per organisation, covers the Computer.

Sources

GitHub, openai/codex#10450: title, closed as completed on 2026-05-14. Checked 2026-11-06.
OpenAI docs, Remote connections: SSH hosts and mobile control. Checked 2026-11-06.
OpenAI docs, Authentication: device code, API key login, auth.json, API keys for automation. Checked 2026-11-06.
OpenAI docs, Codex CLI: standalone installer. Checked 2026-11-06.
GitHub, openai/codex README and source: npm install -g @openai/codex, sign-in host auth.openai.com and ChatGPT backend chatgpt.com. Checked 2026-11-06.