Codex Cloud environments: publish your repository once and tasks keep running with the laptop closed
Codex Cloud environments are set up once and each task runs in its own virtual machine. How to publish them, hours saved, limits and a template.
Codex Cloud environments are a feature OpenAI introduced on September 29, 2026 at DevDay: you set up your repository once, publish it, and every new task runs in its own virtual machine even if your laptop is closed. Here is how an environment gets published, how much time it can save you, who can use it and what still doesn't work.
For the short summary of the launch, read the analysis of Codex Cloud with the laptop closed.
What changed with reusable environments
Before, the cloud prepared your project with the help of a cache. Now you ask Codex to prepare it and test it, and you publish it: an environment that is already installed and verified. The time to set up the project gets paid once, not on every task. From there, every task starts in its own virtual machine from that copy and keeps working even if your computer is asleep. You pick it back up later from the browser or from your phone.
How it works inside
A published environment packages four things: your GitHub repository, the dependencies already installed, the environment variables and network secrets. Network secrets are the clever part: the program only sees a placeholder, and a proxy swaps in the real value when the request reaches an approved domain. By default, internet access is limited to the common package registries, and you widen it if you need to. Every task works on its own copy, so what a task changes never rewrites the original environment.
How to publish an environment in five steps
- You connect your GitHub repository.
- Codex reads it, detects which versions your project uses, writes the install script, installs the dependencies and tests that the workflow runs. If it is missing a variable or a key, it asks you.
- You review what it prepared and publish it. The install script is saved, along with a start step that launches the services and confirms everything responds.
- You launch a task, for example, update the dependencies and run the tests.
- You review the changes and the test results before committing anything.
If you don't like something, you edit the environment and new tasks start from the corrected version.
How many hours it can save you
The math is a WAiBOT assumption, not an OpenAI figure. If setting up your project from scratch on every task takes 8 minutes and you launch 5 tasks a week, that is 40 minutes a week and 2,080 a year: almost 35 hours, more than four eight-hour workdays. The math doesn't depend on the model being smarter, but on setting the environment up once.
Three use cases
These are WAiBOT proposals, not official OpenAI cases. What is official is that each task runs isolated and can be picked up from another device.
- Freelance developer with eight client sites. She publishes one environment per repository in one afternoon and on the first of every month launches the same task on all of them: update dependencies, run tests and summarize what changed. At night she approves the clean summaries and looks closely at the ones that failed.
- Team with a pile of small bugs. One bug, one task. Since every task has its own virtual machine, several can run at once without stepping on each other's files. The new command-line interface brings an agents view to watch several tasks together. How many run in parallel depends on your plan's limits: measure it before promising anything.
- Founder on the phone. With the product on GitHub and a published environment, he types from his phone "add validation to this form and run the tests" and reviews the changes on arrival. There is a possible monthly maintenance service per repository here, where the environment is your asset. Test it with a single client first.
Who can use it today and which machine you get
Access is rolling out gradually on ChatGPT Plus and Pro, and on Business, Enterprise, Edu and Healthcare workspaces, so you may not see it yet. On Plus, each task gets 2 virtual processors, 8 GiB of memory and 8 of disk. On Pro, Business or Enterprise it is 4 processors, 16 of memory and 32 of disk. Check current prices at chatgpt.com/pricing.
Cloud or laptop?
The cloud wins when the work repeats, runs long or can be reviewed later: updates, small bugs, tasks you want to launch and forget. Your laptop wins when you need to watch and talk with the agent in real time, or when your project needs a browser or control of the computer, which the cloud doesn't support yet. The rule: what you can describe in one sentence goes to the cloud; what needs a conversation stays local.
Who it's not for yet and what breaks
- If you work on GitLab or on a self-hosted GitHub Enterprise server: there is no support.
- If you need the agent to use a browser or drive the computer: in the cloud that doesn't exist yet.
- If you have a huge project and a Plus plan: 8 GiB of disk fills up fast.
- If you depend on your personal skills or sign in with an API key: it doesn't work in these environments.
Also, a task's saved state can only be recovered for seven days. Network secrets hide the key but don't limit what can be done with it: give the agent a read-only token. What a task does is not written back to the environment, and sharing an environment doesn't share other people's tasks or let them edit the setup.
How to apply it in your business?
The news isn't running code in the cloud, which already existed with automatic setup and caching. The news is that the environment is now something you publish: tested, with a start step that confirms everything responds, and frozen so every task starts identical. A template for your first environment in six steps:
- Pick the repository that generates the most repeated tasks.
- Connect it and let Codex prepare the install.
- Check which variables and secrets it asks for and use read-only keys.
- Ask it to run the tests before publishing.
- Publish and launch a real task, like updating dependencies.
- Measure how long it took and how much usage it spent.
A model task: update the dependencies, run the tests and summarize in five lines what changed and what failed. Measure for a week before promising anything to a client: ten long tasks aren't free, and the rollout didn't change the per-token rates.
Frequently asked questions
What are Codex Cloud environments?
They are copies of your repository, already installed and verified, that you publish once; every new task runs in its own virtual machine from that copy.
Does Codex keep working with the laptop closed?
Yes, according to OpenAI each task runs in the cloud even if your computer is asleep, and you pick it back up from the browser or your phone.
Who can use Codex Cloud environments?
Access is rolling out gradually on ChatGPT Plus, Pro, Business, Enterprise, Edu and Healthcare.
What are the limits?
A task's saved state can only be recovered for seven days, and there is no GitLab or cloud browser support.
Conclusion
Codex Cloud environments turn project setup into something you publish once, and that makes the cloud useful for repeatable work. Tell me in the video comments: what's the task you repeat most often across your repositories? That's your first candidate.