Codex Cloud: Environments, Workspaces, and How It Differs From Local
The short answer
Codex Cloud executes software engineering tasks on OpenAI-managed computers, not on your local machine. You create a reusable cloud environment containing your repositories, build tools, and setup scripts, then launch tasks that execute inside dedicated virtual machine sandboxes.
Codex Cloud vs. Local Codex vs. Remote vs. Legacy
OpenAI now distinguishes four execution workflows. Developers often confuse Cloud with Remote, or expect Cloud to read local editor state. Here is how they actually differ:
| Workflow | Codex Cloud | Codex Local | Codex Remote | Cloud Legacy |
|---|---|---|---|---|
| Execution host | OpenAI-managed cloud virtual machines | Your local computer (CLI / IDE extension) | Your own machine (remotely accessed) | Earlier OpenAI cloud container fleet |
| When computer sleeps | Continues running uninterrupted | Execution halts immediately | Halts if host computer sleeps | Continued running in legacy runner |
| Workspace isolation | Fresh, isolated workspace VM per task | Shares your local repo working tree | Operates inside host working directory | Task-scoped legacy container |
| Files & state accessed | Cloned repositories + setup script files | Open IDE tabs, uncommitted diffs, localhost | Host machine filesystem and active servers | Connected repo integration snapshots |
| Device access | Desktop app, web, mobile (iOS/Android) | Desktop / laptop terminal and IDE | Web or mobile controlling your computer | Web review dashboards & PR comments |
| Setup requirement | Publish reusable environment (desktop/web) | Install CLI (`npm i -g @openai/codex`) | Configure remote gateway / local tunnel | Authorize GitHub or Linear integration |
| Best for | Long-running tasks, heavy tests, mobile handoff | Interactive pair-programming, local debugging | Monitoring your desktop while away | Legacy Code Review & Security Review |
Do not conflate Cloud with Remote
Codex Remote controls a process executing on your own physical computer via a network bridge. If your laptop shuts down, a Remote task dies. Codex Cloud executes on OpenAI-managed virtual machines in OpenAI's data centers; it continues regardless of your laptop's power state.
Cloud Environments vs. Task Workspaces
OpenAI separates project configuration from individual task execution. Understanding this boundary avoids lost edits and unexpected dependencies.
1. The Cloud Environment (Blueprint)
A reusable project setup that you create and publish on the desktop app or web. It defines:
- Connected Git repositories (GitHub or GitLab).
- Setup scripts (e.g., package installation, build commands).
- Environment variables and proxy network secrets.
- Outbound network policies and allowed destination hostnames.
Draft environments carry an Unpublished badge. You must verify and publish an environment before anyone can launch tasks against it.
2. The Task Workspace (Ephemeral VM)
Every time you trigger a task, Codex spins up an isolated workspace derived from the published environment:
- Tasks start clean from the published baseline.
- Workspaces are strictly isolated from one another.
- Uncommitted files in Task A do not exist in Task B.
- To build on prior edits, reopen the same task thread or commit to Git.
When you edit and republish an environment, only new tasks receive the updated setup. Existing active tasks retain their original state.
Cross-Device Continuation & The 7-Day Recovery Boundary
Codex Cloud allows handing off ongoing development across devices, but saved virtual machine states have strict lifecycles.
Moving between desktop, web, and mobile
Environments must be created and published on the ChatGPT desktop app or web browser. Once published, you can launch, steer, and inspect tasks from:
The 7-Day Saved VM State Recovery Rule
OpenAI documents that, by default, a task's saved virtual machine state remains recoverable for up to seven days after you last start a turn or resume the task.
This seven-day boundary concerns recoverable virtual machine state—allowing you to wake an idle container and resume code editing. It is not your conversation-history retention policy. If you let a task sit untouched for more than seven days, the underlying VM container is purged. Always commit and push important working code to a remote branch before shelving a task.
Plan Availability, Workspace Roles & Quota
Codex Cloud is not open to all ChatGPT tiers. Availability depends on subscription level, account rollout status, and enterprise administrator controls.
| Plan / Workspace | Cloud Tasks | Create Environments | Notes & Restrictions | Limits Guide |
|---|---|---|---|---|
| ChatGPT Free | Excluded | No | Codex Cloud is not included. Free users have baseline local CLI and desktop access. | Free limits |
| ChatGPT Go | Excluded | No | Codex Cloud is not included. Go offers expanded chat and local coding access in select regions. | Plan pricing |
| ChatGPT Plus | Included | Yes | Available to eligible Plus accounts, subject to OpenAI rollout status. | Plus limits |
| ChatGPT Pro ($100 / $200 / $500) | Included | Yes | Included across all Pro tiers. Runs against your plan's included Codex allowance. | Pro limits |
| Business & Enterprise | Included | Yes | Included. Workspace administrators control Cloud access, network rules, and member permissions. | Business limits |
| Education & Healthcare | Included | Yes | Supported subject to workspace configuration and administrative policy. | Enterprise pricing |
Seat-Level Restrictions
In managed organizations, administrators control Cloud feature access. Guest seats, K–12 accounts, and Enterprise view-only seats cannot create cloud environments.
How Cloud Draws From Quota
Cloud tasks draw from your plan's standard included allowance and credit pool. See our detailed breakdown of what counts toward Codex usage.
Step-by-Step: Setting Up a Cloud Environment
Setting up an environment establishes the baseline VM image that future coding tasks build upon.
Connect your source control
Authorize OpenAI Codex to access your GitHub or GitLab repositories. In organization workspaces, a workspace admin must first approve the repository connection.
Create the environment draft
In the ChatGPT desktop app or web interface, open Codex > Environments > New Environment. Select the target repositories you want loaded into the container.
Configure dependencies & setup scripts
Codex inspects your codebase for common manifests (like package.json or pyproject.toml). You can provide custom build scripts to install language runtimes, compilers, and test tools.
Configure network access, secrets & proxy rules
By default, outbound network access is restricted. Add required package registries or API domains to your allowed list. Supply credentials as Network secrets (masked at proxy level) or plain Environment variables (readable by scripts).
Verify and Publish
Codex executes the setup script to verify that packages install and tests compile. Once verified, click Publish. The environment is now active and ready for task delegation across web, desktop, and mobile.
Common Setup & Runtime Gotchas
Here are the most frequent points where developers hit errors when configuring or executing Cloud tasks:
1. "Unpublished" environment badge
If an environment shows an Unpublished badge in Settings, you cannot launch tasks against it. Complete the verification test run in the desktop or web UI and click Publish.
2. Package download or HTTPS errors
Cloud task VMs route traffic through secure proxy filters. If npm install or pip install fails with network errors, verify that the exact registry host (e.g. packages.example.com) is in your allowed domains and uses HTTPS on port 443.
3. Missing local .env secrets
Cloud VMs do not read uncommitted local dotfiles or your laptop's shell variables. API tokens and private registry keys must be explicitly registered under Environment variables or Network secrets in the environment configuration.
4. Starting a new task instead of resuming
If Task A generated a test suite and you start a brand-new Task B, Task B starts with a clean repo snapshot. To build on uncommitted work, reopen the existing task thread rather than starting a new task.
What happened to Codex Cloud Legacy?
The previous Codex Cloud interface remains accessible as Codex Cloud Legacy during the migration period. It continues to handle automated Code Review, Security Review, and existing GitHub or Linear integration triggers.
OpenAI documents that supported legacy environments can be migrated directly into the new environment experience while leaving your original legacy configurations untouched. Past task runs and review history from legacy sessions remain viewable.
Frequently Asked Questions
Direct answers to specific operational questions about OpenAI Codex Cloud.
What is Codex Cloud?
Codex Cloud is an execution model for OpenAI Codex where coding tasks run entirely on OpenAI-managed computers instead of your local machine. You define reusable project environments with your repositories and dependencies, then delegate tasks that execute in isolated virtual machine containers.
Does Codex Cloud keep running if my computer goes to sleep?
Yes. Because Codex Cloud tasks run on OpenAI-managed infrastructure, they continue executing even if your laptop goes to sleep, closes, or loses internet access. You can reopen the task later from the web, the desktop app, or your phone to inspect the results.
What is the difference between Codex Cloud and Codex Remote?
Codex Cloud executes tasks on remote, OpenAI-managed virtual machines. Codex Remote is remote access to a task executing on your own computer. If your computer goes to sleep during a Remote session, the task halts because your local hardware is running the process.
What is the 7-day VM state recovery window?
By default, OpenAI keeps saved VM state recoverable for up to seven days after the last start of a turn or task resume. This recovery window applies specifically to the virtual machine state—allowing you to resume an idle task container—not to your conversation history retention period.
Can I use Codex Cloud on mobile?
Yes. While environments must be created and published on the desktop app or web, published environments can be used to launch, monitor, and continue tasks from the ChatGPT mobile app on iOS and Android.
Is Codex Cloud included with ChatGPT Free?
No. Codex Cloud is not included with ChatGPT Free or ChatGPT Go. Cloud execution requires an eligible Plus account, a Pro tier, or a Business, Enterprise, Healthcare, or Education workspace.
Does Codex Cloud have its own separate quota pool?
No. OpenAI documents that Codex Cloud tasks draw from the same shared usage allowance and credit pool as your local Codex CLI. Multi-step cloud tasks with extensive test runs consume your active rolling usage allowance alongside local turns.
Why can't a new Cloud task see uncommitted changes from an earlier task?
Every new Cloud task starts in its own isolated workspace based on the published environment blueprint. It does not pull over uncommitted files or scratch edits from other tasks. To continue work on uncommitted files, reopen the existing task thread rather than starting a new one.
What happens to active tasks when I edit and republish an environment?
When you edit and republish an environment, the new setup applies only to new tasks launched afterward. Existing active or saved tasks retain their original working state and do not dynamically inherit the environment update.
Who can create Cloud environments in team workspaces?
In managed Business and Enterprise workspaces, administrators control who can access Cloud features. Guest seats, K–12 accounts, and Enterprise view-only seats cannot create cloud environments.
Related Quota & Setup References
Official OpenAI References: