Codex cloud, or Codex on a machine of your own: what each keeps

Two things called Codex cloud

OpenAI’s docs describe two generations under one name, and it helps to know which you are reading about. The older one, now labelled Codex Cloud (Legacy), runs each task in a container from a stock image called universal, caches the container “for up to 12 hours,” blocks internet access during the agent phase by default, and still carries Code Review and the GitHub and Linear integrations, with GitLab in beta on every plan; the docs say “we plan to deprecate this experience.” The current one, announced at DevDay on September 29, 2026, is built on cloud environments: “the reusable setup that tasks use: repositories, dependencies, tools, and access settings. Codex inspects your repositories, prepares the setup, and tests it with you. Each new task gets its own isolated workspace from the published environment.” The rest of this page is about the current one.

A task’s machine

“Each cloud task runs in a VM” whose size follows your plan: on Plus, 2 vCPUs, 8 GiB of memory and 8 GiB of disk; on Pro, Business and Enterprise, 4 vCPUs, 16 GiB and 32 GiB, with larger machines for Enterprise priced on request. The code comes from GitHub, connected when you first set an environment up; GitLab and self-hosted GitHub Enterprise Server are listed as current limitations, and so is “computer and browser use.” Codex writes the install script and the start skill (“instructions to start services and check that they’re ready”) from what it finds in the repository, you review the setup report and select Publish, and “skills stored in your repository are available in cloud tasks. Personal skills from your local computer aren’t synced to cloud environments.” Environments are created on the web or in the desktop app; the phone starts tasks from one already published.

What stays, and for how long

Three kinds of state, which the docs keep apart. A new task “starts from the published environment’s prepared filesystem,” so whatever the install script put there is ready without reinstalling. An existing task “continues with its own saved files, including uncommitted changes and installed tools,” and “by default, a task’s saved VM state is recoverable for up to seven days after you last start a turn or resume the task.” And the environment itself only changes when you republish it: “File changes in a task don’t update the reusable environment,” so a package one task installed is that task’s alone until you edit the setup. Services are the start skill’s job, and nothing documented keeps one running between tasks. The docs put the consequence plainly: “Commit important work or save the output you need. Saved state doesn’t replace source control.” How long a task may run, when an idle one stops, and how many you can run at once are not stated on any page we read; the one figure is the seven days.

What a task can reach

Internet access is a switch on the environment, with presets: package managers (a fixed list of registries and the GitHub hosts), custom domains, or all. “Allowing a destination doesn’t supply credentials or grant permissions in that service,” so credentials go in as environment variables, or as network secrets, where “programs receive a placeholder; the proxy substitutes the real value for allowed destinations.” Internal APIs, package registries and other HTTP or HTTPS services on your own network are reached through a VPN, and “Tailscale is currently the supported VPN provider”; the docs describe that route for HTTP and HTTPS, not for a database port. A public service that filters by source address can allow the published egress addresses, and Enterprise workspaces can ask for OIDC to obtain short-lived cloud credentials. Nothing in the docs gives you a shell into the machine, opens a port on it, or hands out an address for a service running there.

What it costs

The pricing page lists Codex on the web for Plus at $20 a month and above; Pro is $100, $200 or $500; Business is $20 per user a month for two or more seats billed annually, $25 monthly; and an API key gets “no cloud-based features.” The default machine has no price of its own: “local messages and cloud chats share your plan’s usage allowance. Weekly limits may also apply,” and “cloud tasks may use more of your allowance than local messages.” Cloud tasks have counted toward Codex usage since October 20, 2025. When the allowance runs out, Plus and Pro users can buy additional credits, Business, Edu and Enterprise workspaces on flexible pricing buy workspace credits, and anyone can run local chats on an API key at standard rates, which doesn’t help a cloud task. Which plans include which Codex surfaces is its own guide.

From the terminal

The CLI reaches the cloud from your own machine without moving a session there, through a command the docs mark experimental: codex cloud opens a picker of your cloud chats, codex cloud exec --env <id> “submits a task directly” to a published environment, and the stable codex apply applies “the most recent diff from a Codex cloud chat to your local repository,” printing the patched files and failing if git apply hits a conflict. “Continue on web or mobile” means reopening the same cloud task from another device. The desktop app’s handoff moves a chat between your computer and a connected remote host, and the IDE extension can switch a chat into cloud mode; what the remote-connections page rules out is “handoff to a Codex cloud environment,” so a chat running in the CLI on your own computer is not lifted into the cloud.

Where OpenAI puts a machine of your own

OpenAI’s own account of running Codex somewhere other than its cloud is remote connections (the Codex entry in the phone app, still called Remote on older builds): the host is “the ChatGPT desktop app on macOS and Windows,” your phone connects to it, and “if that computer sleeps, loses network access, or closes the app, remote access stops until it’s available again.” Its advice for continuous work is to “add an always-on computer or SSH host when you need continuous access or a different environment”: a dedicated Mac or PC that never sleeps, or an SSH host the desktop app connects to first, after which “remote project chats run commands, read files, and write changes on the remote host.” A Linux machine of your own appears in the docs as that SSH host, as the computer the CLI itself runs on, which is how Codex on a Linux server works, as a CI runner, and, in preview, as a host for the desktop app; the hosts a phone reaches are still the Mac and Windows apps.

When a machine of your own fits better

OpenAI lists when to use its cloud: when “work should run remotely,” “several tasks need the same setup,” “your workflow needs service credentials,” or you want to “start cloud work from the CLI.” The other side of each line is the case for your own machine. A service that has to stay up between tasks, a database, a dev server, a bot, a browser that is signed in, a process that runs for hours: the cloud starts them per task and reaches none of them. Tools and configuration you installed once and want every time: the cloud has what the install script put there and what one task added, for seven days. Work that isn’t on GitHub, or more machine than 4 cores and 16 GiB, or a task you would rather watch in a terminal than read as a diff. And a machine you keep: not state recoverable for a week, but the same computer next month. OpenAI does sell a hosted computer that keeps its state, but to an agent rather than to you: a dot’s computer, on Pro, Business Premium and Enterprise, “has its own files, software, and browser sessions” and is its own thing, not a place to run the Codex CLI.

The condition is the one OpenAI attaches to its own remote connections: the machine has to stay on. A laptop sleeps; a desktop that never sleeps, a server, or a cloud computer of your own does not, and the Codex CLI on it, signed in with your ChatGPT plan and left running in tmux, is reachable from wherever you are over SSH. Everpod’s developer pod is that machine ready made: a cloud computer that is yours, always on, with Claude Code, Codex or both installed, reached only through your own private network, from $24 a month. 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.

What OpenAI says about your code

The pricing page’s plan matrix marks “no training on API or business data by default” as available on Business, Enterprise and the API and unavailable on Plus and Pro, and “enterprise retention and residency controls” as Enterprise only; the finer terms are on OpenAI’s help pages, which refuse automated readers, so read them in a browser before you put private code through a cloud task. For a Business, Enterprise or Edu workspace, the security page for ChatGPT Work’s hosted tasks says execution state and snapshots “follow a separate lifecycle from conversations and files,” governed by the workspace’s retention setting, and that deleting a chat “doesn’t immediately purge every related artifact.” On a machine of your own the files stay on it, but “local execution does not mean offline or device-only model inference”: what Codex reads of them, your prompts and the tool results go to OpenAI to do the work, under the same plan terms.