For development teams
Codex for your company
with shared rules
Your company works with ChatGPT, and your developers have found Codex. Everyone works with their own settings. We set up rules, permissions and the first workflow to a pull request for the whole team.
- ›Take the next item from TODO.md.
- Reads AGENTS.md and the skills
- Sandbox: this folder only, no network
- Tests passed
- Codex review found nothing
- Pull request opened, waiting for approval
What is Codex?
Codex is OpenAI's AI agent for software development, generally available since October 2025. It reads the codebase, changes files, runs commands and tests, and opens pull requests. Codex runs in the terminal, in your IDE, in the ChatGPT desktop app, in OpenAI's cloud and on GitHub. It is part of the ChatGPT plans, including Business and Enterprise. For a team, it pays off once you build software and already work with ChatGPT.
Where Codex runs. Terminal, IDE, app and cloud.
Codex is not one program but one agent with five surfaces. Each has a different place where the permissions are set. Teams that miss this lock down the terminal and forget the cloud.
| Surface | What it is for | What applies there |
|---|---|---|
| Terminal (Codex CLI)on the computer | For developers who work in the terminal, and for scripts: with codex exec, Codex also runs in CI jobs. | The operating-system sandbox, the approval rules and IT's policies. |
| IDEVS Code, Cursor, Windsurf, JetBrains, Xcode | Codex works next to your code, with the open files as context. Changes show up as a diff in the editor. | The same sandbox and policies as in the terminal. |
| ChatGPT desktop appmacOS, Windows, Linux | Several agents work on tasks in parallel, with Git built in. People who don't want a terminal work here. | IT's policies for every machine and the approvals in the workspace. |
| Codex Cloudon OpenAI's servers | Each task keeps running in its own workspace, even while your computer is asleep. It ends in a commit or a pull request. | The workspace admin. In Enterprise, the cloud stays off until the admin turns it on. Repository permissions in GitHub or GitLab decide which repositories it can see. |
| GitHub and Slackin the pull request and the channel | With @codex review, Codex reviews pull requests, automatically on every one if you want. In Slack, @ChatGPT hands tasks to Codex Cloud. | Whoever connects the repository and approves the Slack app. For Slack, that is a workspace owner or admin. |
Everyone on the team uses Codex differently.
Codex comes with the ChatGPT plans. Teams often work with it before anyone has set rules for it.
- When something works for one developer, the next one never hears about it.
- Nobody has decided whether Codex may access the network or which folders it may change.
- Pull requests from the agent reach the main branch without a second look.
AGENTS.md, Team Config, skills: the rules live in the code.
Codex reads an AGENTS.md before it starts working. It lives in the repository, and the team maintains it like any other code. A good AGENTS.md is short and says three things: how to build and test, which rules apply, and where the agent never touches anything. What does not belong in it: long explanations, knowledge the code already holds, and credentials. AGENTS.md is an instruction, not a lock. What must really stay closed is set by the sandbox in the next section.
# AGENTS.md
## Build and tests
- Dependencies: npm ci
- Before every commit: npm run check and npm test
- A pull request is only done when CI is green.
## Rules
- New features only with a test.
- No new dependencies without asking.
- Never read or change .env, secrets/ or infra/.
- Deploy only through a pull request, never by hand.AGENTS.md on two levels
One file applies to each user, one per project. The file closer to the working folder takes precedence. By default, Codex reads up to 32 KiB of it, and a long file gets cut off. Claude Code, Cursor and GitHub Copilot read the same file, so switching tools later does not start from zero.
Team Config under .codex
Defaults and rules for commands outside the sandbox live in the repository. The whole team works with the same setup instead of everyone configuring their own.
Skills under .agents/skills
A skill is a folder with a SKILL.md that follows the open Agent Skills standard, for example for “prepare a release” or “create a migration”. A workflow then runs the same way every time. Skills you wrote for Claude Code also run in Codex once they sit in .agents/skills or are linked there.
Sandbox and approvals: how Codex works safely.
On a computer, Codex works inside an operating-system sandbox. By default, it writes only in the current working folder and has no network access. If it wants to change files outside it or run a command that needs the network, it asks. Your team and your IT agree on these limits together, on three levels.
Set centrally
Which permission profiles and approval rules are allowed, whether browser and computer control stay off, which tools agents use and how they access files and the network, locally and in Codex Cloud. This runs through requirements.toml and through Agent Security in the workspace admin area.
The team may loosen per project
Within those limits: allow single commands without asking, such as the tests, or open network access to one package server for one project. This sits in the Team Config and is visible to everyone in the repository.
Always stays closed
Production credentials, write access outside the repository, and actions that delete data. A person always approves tool calls marked as destructive. In Codex Cloud, programs only see a placeholder, and the proxy inserts the real key for approved destinations.
This is what IT's policy looks like. The file reaches every machine through device management or the workspace, and no project and no developer can override what it says.
# requirements.toml, delivered by IT
default_permissions = ":workspace"
allowed_approval_policies = ["on-request"]
[allowed_permission_profiles]
":read-only" = true
":workspace" = true
[features]
browser_use = false
computer_use = falseHow a task moves through the team with Codex.
Codex works best on tasks with a clear goal and tests. This workflow runs in our own operation, and we set it up for you the same way.
Issue
Goal, non-goals and acceptance criteria are in the ticket before the agent starts. The ticket is the brief.
Agent run
Codex works locally or in the cloud, several tasks in parallel. Each task runs in its own workspace.
Tests in CI
Type checks and tests must pass. Red goes back to the agent with the error log.
Codex review
@codex review reads every pull request and comments on findings before a person looks.
Approval by a person
A developer checks the result and brings it into the main branch. That approval stays with a person.
What an issue Codex can work with looks like
Three fields are enough. Anything you leave out, the agent guesses.
Goal: the export button in the customer list
returns a CSV with all visible columns.
Non-goals: no new design, no new
dependencies.
Acceptance: the export test passes,
umlauts are correct, the file opens in Excel.Good fit
- Add missing tests
- Reproduce and fix a reported bug
- Prepare a migration that follows a fixed pattern
- Update documentation after a change
- Update dependencies when tests cover them
Poor fit
- Architecture decisions
- An application's security model
- A large rebuild without tests
- Tasks where nobody can say when they are done
Where the code goes, and what the plan controls.
Codex sends instructions and the code excerpts it needs for the task to OpenAI's models. That applies in the terminal and in the cloud. In Codex Cloud, a copy of the repository also sits in the task's workspace.
That is why Codex belongs in a company plan. With private accounts, you control neither the settings nor the contract. On the Business plan, OpenAI does not train on your business data by default.
Credentials do not belong in the prompt or in the AGENTS.md. Locally, they live in the password manager; in Codex Cloud, as secret environment variables that are only inserted for approved destinations. Codex can read a .env file in the working folder, so it does not belong in the repository. Don't connect repositories with especially sensitive data to the cloud at all.
| Plan | What IT controls with it |
|---|---|
| ChatGPT Business | A dedicated workspace with admin settings, single sign-on through SAML and multi-factor sign-in. |
| ChatGPT Enterprise | Plus user provisioning through SCIM, your own encryption keys, role-based permissions, usage reports, audit data through the Compliance API, and settings for data retention and data residency. |
It depends on where your company already works.
All three read an AGENTS.md, and each needs rules, permissions and a review step. What tips the balance is which systems your team uses today.
When Codex fits
Your company already works with ChatGPT Business or Enterprise. Development then needs no second account and no second contract, and access and permissions live in the same workspace.
When Claude Code fits
Your company's business teams already work with Claude, and development should use the same skills and integrations.
When GitHub Copilot fits
Your code lives on GitHub, and you want the agent, code review and licenses in the place where your repositories, issues and pull requests already are.
Four steps to the first workflow.
Week 1: plan and access
With you and your IT, we settle the plan, who uses Codex locally and who in the cloud, and sign-in through your company account. Result: everyone on the team works with the company account.
Weeks 1 to 2: policies for every machine
Your IT rolls out requirements.toml and sets Agent Security in the workspace. Result: sandbox, network access and disabled features are the same on every machine.
Week 2: rules and skills in the repository
AGENTS.md, Team Config and the first skills go into the repository, with templates from our own operation. Result: every session starts with your rules.
From week 2: the first workflow
Codex review runs on every pull request, and a first skill takes an item from your backlog to a pull request. Result: a workflow that runs every week. Then we move on use case by use case.
What Codex does not solve.
No architecture
Codex builds what the issue says. How the system should be designed is still your team's call.
Models get retired
OpenAI regularly removes older models from Codex. If a model is hard-coded in scripts or policies, the workflow stops. That is why we keep the model choice in one place.
The network block slows things down
Without network access, some builds don't finish. We open single package servers instead of the whole network.
Cloud tasks need a setup
In Codex Cloud, the environment has to know how the project is built. We set this up once per repository.
Quality depends on the issue
A vague ticket gets a vague pull request. Writing down the goal and acceptance criteria first saves the most time.
Usage needs watching
Every task counts against your plan's allowance, large cloud tasks more than small ones. We review usage with you once a month and adjust the model choice.
How we use Codex ourselves.
Our main tool is Claude Code. Codex runs next to it, with the same rules and the same skills.
Frequently asked questions.
Which is better, Codex or Claude Code?
Neither is better; the two agents can do almost the same things. What tips the balance is whether your company works with ChatGPT or with Claude. The detailed comparison with GitHub Copilot and Cursor is on our Claude Code page.
What can Codex do?
Codex reads a codebase and explains it, writes and changes code, runs commands and tests, fixes bugs, adds tests and opens pull requests. It reviews pull requests on GitHub, runs in scripts and CI jobs, and works on several tasks in parallel in the cloud. With skills, it learns your team's fixed workflows.
What do companies need to consider when using Codex?
Three things: a company plan instead of private accounts, central policies for sandbox and network access, and a review step before code reaches the main branch. Add a data processing agreement with OpenAI and a decision on which repositories Codex may see. Whether that is enough for your data is for your data protection officer to judge.
Can I use Codex for free, and which plan fits a team?
Codex comes with the ChatGPT plans, and for a team that means Business or Enterprise. How to roll out ChatGPT Business for every department, with workspace, data protection and apps, is on our ChatGPT Business page.
Can we switch from Claude Code or Cursor to Codex without starting over?
Yes. The ChatGPT desktop app imports instructions, settings, skills and plugins from Claude Code, Claude Cowork and Cursor, and the terminal client imports from Claude Code and Cursor. The old setup stays as it is.
Where is your team stuck with Codex?
Tell us what you have tried so far. In 30 free minutes, you get a clear view of what is doable for you and what we would advise against.
