Codex Cloud

Give Codex a failing test before leaving your desk and let the task keep running while your laptop is off. Codex Cloud gives you a reusable project environment, with web and phone access to follow the work, steer it, and review the result.

What Changed on September 30

The relaunch makes cloud work easier to repeat: your repository, dependencies, scripts, and settings can be ready before the next task starts. You can follow progress and steer Codex from your phone or another computer. The familiar desktop experience also comes to web and mobile. Those are the changes in OpenAI's September 30, 2026 announcement.

Think of the environment as a prepared workshop. The tools and materials stay ready; each task gets its own workbench. A dependency is a software package your project needs, and the repository is the versioned collection of its code.

The practical payoff is separating coding work from keeping your own machine awake. Codex Cloud runs on OpenAI-managed computers. Create and publish the environment on desktop or web, then use it across supported devices. Each task has its own working files. OpenAI's Cloud help guide explains that separation.

My recommendation: start with a small job whose result you can verify. A background agent is most useful when you have already decided what a correct outcome looks like.

Which Plans Include Codex Cloud?

Cloud access requires an eligible Plus, Pro, Business, Enterprise, Healthcare, or Education account, subject to rollout and workspace settings. Free and Go include Codex access, but do not include Codex Cloud. Guest, K-12, and Enterprise view-only seats cannot create cloud environments. OpenAI's plan availability is the source for these distinctions.

Subscription prices above come from the current Codex pricing page, checked October 7, 2026. Cloud eligibility does not mean every member can use every control.

How Cloud Work Uses Your Allowance

At launch, standard environments have no separate virtual machine charge. A virtual machine is the remote computer running the task. Model usage still counts toward normal Codex limits and applicable credits or billing. Plans with shared pools can share usage across Codex, ChatGPT Work, ChatGPT for Excel, and Workspace Agents when available; token-based Enterprise agreements bill in USD instead of credits. OpenAI's usage rules spell out that accounting.

Local messages and cloud chats share your plan's allowance, and weekly limits may apply. Cloud tasks can consume more allowance than local messages. The pricing page's message estimates describe local usage, not a guaranteed number of cloud jobs. Eligible Plus and Pro users can buy extra credits without upgrading. Check your usage dashboard for current limits and reset times. Codex pricing and usage explains the variables.

If you already pay for an eligible plan, try a bounded task before buying more capacity. The business math is hands-on time avoided minus setup maintenance, steering, and review time, with any additional credits counted separately. Moving a task off your laptop only pays if that balance improves.

For the wider subscription and API comparison, see Codex Pricing in 2026. Cloud requires ChatGPT sign-in; a local CLI session using an API key follows separate API billing. Authentication options explain the difference.

Set Up One Reusable Environment

Prepare a project whose tests run reliably before delegating a larger change. Follow the current cloud environment setup, rather than the older legacy workflow.

1. Connect the repository. On web or desktop, choose Work in > Cloud > Select environment > Create environment. Select GitHub repositories and connect GitHub if prompted.
2. Prepare the project. Select Get started. Let Codex inspect, install, and test. Supply missing information and required versions.
3. Review the setup script. Install script records dependency preparation; Start skill records service startup and readiness. Refine setup conversationally.
4. Configure values. Select Manage beside environment variables or network secrets. For a network secret, enter its key, value, and allowed domains.
5. Set internet access. Enable Allow Codex to access internet if needed. Choose Package managers or Custom domains only, adding required hosts. All (unrestricted) permits broader access.
6. Publish and start. Review files, configuration, and checks. Save, then select Publish. After Environment published, start a new task. Later setup changes use Edit and Republish.

For the preparation conversation, give Codex the project's actual install and test commands, required runtime version, and services. Ask it to report what passed and what it could not verify. These are suggested instructions, not a universal setup script.

For a first experiment, I would use synthetic test data and the narrowest service access that reproduces the job. If a package download fails, inspect its hostname and authentication separately instead of immediately widening internet access.

Three Good First Tasks

Choose a task with a clear starting point, a small scope, and evidence you can inspect afterward. These task briefs are suggestions you can adapt to your repository.

Fix a Failing Test

For a founder whose release is blocked by a reproducible failure, delegate the investigation and a focused patch. Supply the failing command and output. Ask Codex to preserve the intended behavior, explain the cause, and report the checks it ran.

A useful brief:

Reproduce this failing test using the project's documented command. Find the cause and make the smallest correct fix. Do not weaken the assertion to make it pass. Run the focused test and relevant nearby tests. Summarize the changed files, results, and remaining uncertainty.

This is my strongest first task because it has a visible before-and-after condition. A green result still needs a diff review: the patch must fix the behavior, not merely hide the failure.

Write a Migration

For a backend developer changing a database schema, ask for the migration file, compatibility considerations, and tests against disposable data. A migration is a versioned change to the database structure or stored data.

A useful brief:

Draft the migration for this schema change using the repository's existing conventions. Explain compatibility with the current application, rollback options, and data-loss risks. Test against disposable fixtures where possible. Do not run it against production or deploy it.

The payoff is a reviewable implementation and a clearer rollout plan. Writing a migration does not settle production locking, backfill duration, or whether the old and new application versions can coexist. Keep those decisions with the person responsible for deployment.

Review a Pull Request

For a maintainer waiting on a teammate's changes, identify the PR's branch or commits and its base branch. Ask for findings with evidence, while keeping the working files unchanged.

A useful brief:

Inspect this pull request against its base branch. Focus on correctness, authorization, compatibility, and missing tests. Give each finding a file location, concrete failure scenario, and supporting evidence. Do not change files or merge the PR.

If you want the integrated GitHub review, OpenAI documents a connected repository and an @codex review PR comment. GitHub review setup covers that path. Code Review, Security Review, and existing GitHub and Linear integrations continue to use Codex Cloud (Legacy) during the transition. The Cloud help page confirms this distinction.

A review request inside your new cloud task and an automatic GitHub review are separate workflows.

Three More Jobs Worth Delegating

After a failing-test fix, I would rank these next by how easy they are to scope and verify. Each is a proposed workflow, not a reported result.

Keep the acceptance condition in the brief. “Improve this codebase” leaves too many decisions unresolved to be a useful first assignment.

Follow and Steer From Your Phone

Return to the same task when you want to continue its work. Opening a new task starts a separate workspace and does not recover the first task's uncommitted changes. Commit important work. The default saved VM recovery window is up to seven days after the last turn start or task resume, rather than a conversation-history retention rule. OpenAI's task-state guidance explains what is saved.

From mobile, open Codex and choose the published environment. Reopen the task to follow progress and send corrections. The Cloud overview describes the cross-device workflow.

Good steering messages resolve a decision:

• “Keep the fix in the parser; preserve the public response shape.”
• “Use a disposable database fixture for the migration test.”
• “Stop after the patch and test report; leave deployment for review.”

I would use the phone to resolve scope and check progress, then review a substantial diff on a larger screen. Remote access to a task executing on your laptop still depends on that computer; it does not provide Cloud's laptop-off execution. OpenAI's help page distinguishes the two.

Cloud or Local CLI: Which Wins?

Use cloud when a well-scoped job should continue independently. Use local CLI when the task depends on the files and developer tools already on your machine.

The CLI can inspect a local repository, edit files, and run installed tools. Open the project directory, run codex, and sign in with ChatGPT. It can also delegate through codex cloud, so the interface you start from does not determine where the work executes. The CLI guide covers both.

The choices below are my recommendations.

Current Cloud environments do not support computer/browser use, GitLab, or self-hosted GitHub Enterprise Server. Personal local skills are not synced. Current limitations matter more than a general preference for cloud.

Keep Access Narrow and Review the Diff

Know what the environment can reach. Allowed domains govern VM network destinations; they do not grant service permissions. Environment-owned network secrets also allow their domains. Enterprise Agent Security requirements apply alongside environment settings. Network configuration and Agent Security describe those controls.

Choose the right credential delivery. Direct environment variables reach programs. Network secrets use proxy placeholders for approved HTTPS destinations on port 443 during setup and tasks, keeping raw credentials out of local processes and files. Secret handling explains the mechanism.

I would keep production credentials out of a first coding task and use narrowly scoped development access. Before merging, read the diff, check test output, inspect dependency changes, and confirm that the patch matches the brief. A passing test suite is evidence to assess, not permission to skip review.

Healthcare account eligibility does not make Cloud covered by OpenAI's BAA, its agreement for covered health-data handling. OpenAI says not to process protected health information in Codex Cloud. Cloud data restrictions state that boundary.

For team controls, see How to Set Up Codex Security After DevDay.

What You Could Build Around This

The strongest opportunity is a release-blocker repair kit for one software stack. Sell the repeatable workflow and validation to teams that lose time on broken tests. DataForSEO's October 7 check estimates 1,600 US monthly Google searches for “automated software testing tools.” That measures interest in the job, not demand specifically for Codex Cloud.

The smallest useful version could be repository instructions, reproducible fixtures, focused repair briefs, and an environment preparation guide. Measure whether proposed patches preserve intended behavior and reduce review effort. The catch is varied test infrastructure: broad stack support would make a small kit expensive to maintain.

A migration verification pack could serve teams using one framework and database. DataForSEO estimates 1,300 US monthly searches for “database migration tools.” An MVP could supply migration templates, disposable test data, compatibility checks, and review prompts that a developer runs in a prepared environment. The catch is production behavior: a reusable package cannot promise that a real database rollout will be safe or quick.

A PR review evidence pack could help maintainers standardize how they ask for and assess findings. DataForSEO estimates 1,300 US monthly searches for “ai code review,” and the live search questions include “Can ChatGPT do a code review?” Start with repository guidance, review briefs, and an evidence format for findings. The catch is that native review already exists; the package must add domain-specific judgment and useful checks.

All three are proposed products built around documented task delegation. The search numbers are estimated job-level interest, not customer counts or revenue forecasts. Start with the repair kit: its acceptance condition is easier to measure than the general quality of a review.

Can ChatGPT do a code review?

Yes. Codex has a documented GitHub PR review workflow, and you can also request an inspection task. Specify the base branch and the risks you want checked. Keep the human merge decision.

Is the AI code safe?

Assess the actual patch. Review changed behavior, authorization, dependencies, tests, and any unverified assumptions. Neither a polished explanation nor green tests establishes that every important case is covered.

Are code reviews worth it?

Use an agent review where another inspection could catch an expensive mistake. Track useful findings and false alarms. If it creates more review work than it saves, narrow the scope.

Your Monday Move

Pick one repository with a reliable test command. Prepare and publish its environment, delegate a small reproducible failure, and check the task from your phone after leaving your desk. Review the patch before merge. Record setup effort, review effort, and plan usage. Reuse the environment if that experiment improves your workflow.

If you want this workflow built into your team's delivery process, build an AI production system.