How to use Codex app for non developers: setup, plans and a safe first task

Is the Codex app a desktop download or a browser tool?

Every platform, plan and permission detail on this page was checked against OpenAI's own documentation on 19 October 2026.

Codex is a mode inside the ChatGPT desktop app, and the web version does a narrower job. The OpenAI quickstart says "The ChatGPT desktop app is available for macOS, Windows, and Linux." The Linux build is a preview: OpenAI's Linux install page lists supported versions of Ubuntu, Debian, Fedora and Arch Linux, each with x64 and ARM64 packages.

The browser side is Codex Cloud. The Codex Cloud page describes starting tasks on the web, mobile or desktop from a published cloud environment built from your GitHub repositories. That suits a team with a repository. It does not suit a first session on a folder of spreadsheets on your laptop, because a cloud task never sees that folder.

For a non-developer, the decision is simple: install the desktop app and start locally.

Which ChatGPT plans include Codex?

All of them, with limits that differ by plan. OpenAI's pricing page states that "ChatGPT Work and Codex are included in your ChatGPT Free, Go, Plus, Pro, Business, Edu, or Enterprise plan". The detail underneath matters more than the headline:

Plan	Codex in the desktop app	Codex cloud tasks on the web
Free	Listed as subject to rollout	Not listed
Go	Listed as subject to rollout	Not listed
Plus	Yes	Yes
Pro	Yes	Yes
Business	Yes	Yes
Enterprise and Edu	Yes	Yes

Usage is the real constraint. For Plus, the pricing page estimates 15-160 local messages per five-hour period with the GPT-6.1 Sol model, and 350-3,000 with GPT-6 Luna. OpenAI calls these estimates, not fixed limits, and a long task on a big folder uses far more per message than a small one, so the range tells you the order of magnitude, not how many tasks you get. It still shows how much further GPT-6 Luna stretches a Plus allowance than GPT-6.1 Sol does. Plans and limits change often; check that page before you upgrade for Codex alone.

Should you pick Codex or ChatGPT Work for a task?

Pick Codex when the output is something that runs, and ChatGPT Work when the output is something you read. The quickstart puts it plainly: "For software development with codebase context and developer tools, select Codex from the ChatGPT dropdown." The same page sends research, analysis, documents, presentations, spreadsheets and Sites to ChatGPT Work.

For a product manager, the Codex side of that line covers more than it sounds. A clickable prototype is software. So is a script that merges three CSV exports the same way every Monday, or a one-line copy fix in your team's repository. A competitive summary, a PRD draft or a meeting recap is Work territory, even though Codex could produce it.

How do you set up the Codex app the first time?

Setup takes four steps, all from the desktop app page and the quickstart:

• Download the ChatGPT desktop app from the Download ChatGPT link on OpenAI's desktop app page, or follow the Linux install guide on Linux.
• Open the app and sign in with your ChatGPT account.
• Open the ChatGPT dropdown and select Codex, then open a folder or create a local project. Codex can read and change the files in the folder you choose, so choose a folder that holds only the work for this task.
• Choose where the chat runs. The environments page lists three options: Local works in the folder directly, Worktree isolates changes in a Git worktree, and Cloud runs remotely. "Both Local and Worktree chats run on your computer."

For a first session, make a copy of the files you want to work on and open the copy with a Local chat. Worktree only applies to a folder that is already a Git project, which most non-developers do not have yet, and a copy gives you the same safety without learning Git on day one.

Which permission setting is safe for a non-developer?

Ask for approval, and leave it there for your first weeks. The permissions page tells you to "start with Ask for approval" for most work, and describes it as letting Codex work within the current workspace while it "pauses before reaching beyond that boundary." The other options are Approve for me, which sends those boundary requests to automatic review but keeps the same sandbox, and Full access, which the app describes as "Unrestricted access to the internet and any file on your computer".

OpenAI's Codex best practices say the same to beginners directly: "If you're new to coding agents, start with the default permissions." The page lists "Giving Codex full permission to your computer before you understand the workflow" among the common mistakes. When Codex asks to do something you do not understand, say no and ask it to explain the request in plain language first.

What is a good first task for a non-developer?

A good first task produces a result you can check without reading any code. Cleaning a data export is the classic one: you already know roughly how many rows and which columns it should have. The best practices page recommends a prompt with four parts (goal, context, constraints, and "Done when"), which also forces you to write down how you will check the result:

Goal: combine the three CSV exports in this folder into one clean file.
Context: the files are signups-jan.csv, signups-feb.csv, signups-mar.csv.
Constraints: do not change or delete the original files. Write the result
to a new file called signups-q1-clean.csv. Explain each cleaning rule you
apply in plain language.
Done when: the new file has one header row, no duplicate emails, and you
report the row count for each input file and for the output.

Before Codex touches anything, switch on Plan mode with /plan. The best practices page says "Plan mode lets Codex gather context, ask clarifying questions, and build a stronger plan before implementation", and for a non-developer the plan is the part you can actually review: it lists the rules in words before they become code. Read it, correct any rule that is wrong for your data, then let it run. Afterwards, open the output, check the row counts against what you expected, and spot-check five rows by hand.

When is Codex the wrong tool for a non-developer?

Codex is the wrong tool when you cannot tell whether its output is correct, and that limit does not go away with a better prompt. A cleaned spreadsheet is checkable. A new login flow in a production app is not, for someone who cannot read the diff, and that work belongs with an engineer reviewing it. The guide to when not to use an AI coding agent covers the cases in more depth.

Two practical limits catch people early. Cloud tasks do not see your machine: the environments page says your "local files, running processes, browser sign-ins, and VPN access aren't automatically transferred" to a cloud task. And OpenAI's pricing page says weekly limits may also apply on top of the five-hour estimates, so a recurring job you depend on should be checked against your plan's limits before you build a routine around it. If the vocabulary still feels foreign, what agentic coding means for product managers explains the agent loop without code.

Where do you go after the first week with Codex?

Once two or three file tasks have worked, the next step is a real repository, and the order matters. Start with a short instruction file: how to write an AGENTS.md for Codex gives Codex your project's rules so you stop retyping them. Then the Codex feature build loop shows how to take a small feature from plan to slices to a commit. Reviewing Codex pull requests covers what to check before anything merges, and Codex skills turn a task you repeat into one you name and rerun.

Building with Codex is one of Builders Camp's 40+ bootcamps, and each bootcamp is led by its own programme director.