OpenAI Codex Cloud Environments Guide: Setup, Shared Configurations, Secrets and Security Cloud
OpenAI gave Codex reusable cloud environments at DevDay 2026, so tasks keep running when your laptop is closed and teams can share one approved setup. How environments work, plan limits, secret handling, Codex Security Cloud and the new CLI.
OpenAI’s Codex coding agent has been able to run tasks in the cloud since its research preview in May 2025, but each task used to start from a fresh, isolated sandbox. At DevDay on September 29, 2026, OpenAI introduced reusable cloud environments: saved project configurations with dependencies already installed, which any task can start from and which teams can share. Together with a refreshed CLI, a new code review view and Codex Security Cloud, the update makes Codex less tied to a developer’s laptop. This guide explains how the pieces work and how to set them up sensibly.
What changed at DevDay
As TechCrunch reports, OpenAI’s pitch is that developers can run Codex wherever they need it: on a computer, remotely from a phone, or in the cloud. The reusable environments are designed to help tasks start faster and to give teams a shared workspace with approved settings and permissions.
The other Codex announcements:
- Codex CLI refresh with voice control for starting and steering tasks, an
/agentsview for delegating and tracking several tasks, and improvements to prompt editing, session resuming and Git worktrees. - Code review in the ChatGPT desktop app, where users can read summaries, explore diffs and ask Codex about potential issues before commenting on GitHub pull requests or GitLab merge requests, plus automatic first-pass reviews.
- Codex Security Cloud, covered below.
How cloud environments work
SiliconANGLE’s detailed write-up describes the workflow:
- Connect a GitHub repository. Codex reads the code to identify the software versions the project depends on and drafts an install script.
- Refine the setup in conversation. Startup checks for the project’s services are generated in the same pass, and changes are made by talking to Codex.
- Publish the environment. From then on, each task gets its own virtual machine starting from that saved setup, and keeps working while your computer sleeps.
- Pick up anywhere. Unpushed changes stay with the task, so it can be reopened from a phone or browser. Dependency caches survive the background refreshes that keep the repository current.
By default, a task’s state can be recovered for up to seven days after it was last used.
Sharing with a team
An environment can be shared with a ChatGPT workspace so colleagues can run their own tasks from the same approved configuration. Sharing does not give access to other people’s tasks or the right to edit the setup; enterprise administrators grant editing permission separately. This is useful for onboarding: a new developer can start working from a known-good environment instead of reproducing a local setup.
Secrets and network access
Environments support standard environment variables and a second type OpenAI calls network secrets. A program using a network secret sees only a placeholder; a proxy inserts the real value as traffic goes to an allow-listed domain. That reduces the risk of a secret being printed in logs or exfiltrated by code the agent runs. In shared environments, individual users can be asked to supply their own values.
Internet access defaults to a preset of common package registries, and developers can widen it if needed.
A typical workflow after setup
Once an environment is published, a developer might assign a bug fix from the ChatGPT app on their phone during a commute, let Codex reproduce the issue and run the test suite in its own virtual machine, then review the diff in the desktop app’s code review view later that day. If the change needs more work, the task can be reopened with its unpushed state intact. Meanwhile, the CLI’s /agents view shows every delegated task in one place, which helps when several run in parallel.
Plans and limitations
According to SiliconANGLE, cloud environments are rolling out to ChatGPT Plus and Pro subscribers and to Business, Enterprise, Edu and Healthcare workspaces. Plus accounts get the smallest virtual machines, with half the processors and memory of Pro, Business and Enterprise tiers.
Current limitations:
- No support yet for repositories on GitLab or self-hosted GitHub Enterprise Server.
- Computer and browser use are not supported inside cloud environments.
Codex Security Cloud
Codex Security Cloud scans entire GitHub repositories on demand, on a schedule, or whenever new commits arrive. Codex then investigates findings, removes duplicates and prepares fixes in the cloud, even when the developer’s laptop is closed. The Decoder notes that it tracks each vulnerability from discovery to fix and includes access to OpenAI’s Daybreak Blue defensive security models without a separate application. It is available on Pro, Business, Enterprise and Edu plans, and complements the open-source Codex Security CLI released in July for command-line and CI/CD use.
The Ultrafast tier
The Decoder also reports a new “Ultrafast” tier that speeds up token generation considerably and costs six times the standard rate in the API. It is available for GPT-6 Astra on Enterprise plans and on the new Pro 500 subscription at $500 a month. For most coding work, standard speed is likely to be enough; Ultrafast suits interactive sessions where waiting is the main cost.
Setting up Codex cloud environments well
- Start with one repository that has a clear setup script and tests, and check that Codex’s generated install script matches what your CI does.
- Use network secrets for credentials rather than plain environment variables wherever possible.
- Keep the network allow list narrow, adding domains only when a task needs them.
- Separate test and production credentials. Agent environments should never hold production secrets.
- Require pull requests for all agent changes and keep branch protection on main.
- Share environments for consistency, but restrict who can edit them.
- Pair Security Cloud findings with human review before merging fixes.
How it compares
Codex’s model keeps both the agent and the execution environment in OpenAI’s cloud. That is the simplest option, but some companies prefer to run execution in their own infrastructure, as Cursor now allows with self-hosted machines on Cloudflare Sandboxes. Anthropic’s Claude Code offers another approach with parallel agents in the terminal. Whichever tool you choose, the same security principles apply; see our guide to securing AI coding agent plugins.
Our analysis
Reusable cloud environments solve a practical problem: the setup work that happens before an agent can write useful code. By saving that work once, sharing it across a team and handling secrets through a proxy, OpenAI has made Codex more suitable for real team workflows rather than one-off tasks. The current gaps, especially GitLab and self-hosted GitHub support, will matter for many enterprises. For teams already on GitHub and ChatGPT plans, it is worth piloting on one repository with tight network and secret settings.