Codex, properly understood

1. Step 1 of 16

   Chat, Work or Codex?

   ChatGPT has three modes, and they share one engine:

   • Chat is ordinary back-and-forth: questions, ideas, drafts.
   • Work carries a bigger task through to a finished result, like a report, a slide deck or an analysis, and keeps the technical detail out of sight.
   • Codex is the same capable agent, but it shows you its workbench: the files it changes, the commands it runs, and every edit line by line.

   Work and Codex share a usage allowance. Pick by how much you want to see:

   • Work when you want the finished document.
   • Codex when you are building or changing something made of files and want to check each change.

   Think of it like…

   Chat is phoning a knowledgeable friend. Work is hiring an assistant to hand you a finished report. Codex is the same assistant, but you sit beside them at their workbench.

   You’re done with this step when

   You can say which mode you would use to: ask a question, get a finished slide deck, and build a small website.

   I did this

2. Step 2 of 16

   Open Codex in the ChatGPT app

   1. Download the ChatGPT desktop app from OpenAI if you do not have it, and sign in.
   2. Open the ChatGPT dropdown at the top and choose Codex.
   3. Click New chat.

   You’re done with this step when

   You are looking at a new Codex chat in the desktop app.

   For the curious

   Codex also comes as a terminal tool. On Mac or Linux:

   curl -fsSL https://chatgpt.com/codex/install.sh | sh
   

   Then type codex in a folder and choose Sign in with ChatGPT.

   There is an extension for VS Code, Cursor and Windsurf, and Codex Cloud (Plus and above) runs tasks on OpenAI's servers against GitHub projects.

   I did this

3. Step 3 of 16

   Give it a practice folder

   Codex reads and changes files only inside the folder you choose, so start with a safe one.

   1. Make an empty folder on your Desktop called codex-practice.
   2. In Codex, open that folder.
   3. Under Work in, keep This computer and Local.

   If Codex offers to set up Git (a built-in history of changes), say yes. It is what lets you review and undo.

   You’re done with this step when

   Codex is open on your codex-practice folder, working locally.

   I did this

4. Step 4 of 16

   Make something, then review it

   Ask for something you can see. When Codex finishes, open the review pane:

   • It shows exactly what changed, line by line.
   • Keep what you like, and revert what you do not: for one file, one part, or everything.
   • Last turn shows only what Codex just did.

   Then double-click the new file in your folder to open it in your browser.

   See an example: A vague request vs a clear one ▸

   Try this

   Create a simple one-page website (index.html) for a bakery: its name, opening hours, a small menu table and a contact section. Make it look good on a phone.

   Copy

   You’re done with this step when

   You opened the review pane, looked at the change, and opened the page in your browser.

   I did this

1. Step 5 of 16

   The context window: what Codex can see right now

   Codex works only with what is in its context window: the maximum amount it can consider at once. That includes:

   • your messages
   • its standing instructions
   • every file it reads
   • the output of every command
   • descriptions of the tools and skills it has

   Context is counted in tokens, chunks of text of roughly three-quarters of a word. The current models can hold about a million, but a long task that reads big folders still fills it.

   This is why Codex can lose track of something from early in a long chat.

   Think of it like…

   A whiteboard. Everything it is thinking about must be written there, and when it is full, something gets wiped or summarised.

   You’re done with this step when

   You can explain why a very long chat might lose details from its start.

   For the curious

   In the terminal, /status shows how much context is left, alongside your rate limits.

   The docs' guidance is to keep one chat per piece of work and to push noisy exploration into subagents, avoiding what they call "context rot".

   I did this

2. Step 6 of 16

   When it fills up: compaction and fresh starts

   As the window nears its limit, Codex compacts: it summarises the older part of the conversation so work can continue. A summary keeps the gist but loses detail.

   Your controls:

   • /compact summarises now, before a big next step.
   • /new starts a fresh chat.

   The habit that saves most frustration: one task per chat. Finished the website? Start a new chat for the spreadsheet.

   Think of it like…

   Photographing a full whiteboard, writing a summary in the corner, and wiping the rest.

   Try this

   /compact

   Copy

   You’re done with this step when

   You know when you would compact and when you would start a new chat.

   I did this

3. Step 7 of 16

   History: resume, fork and archive

   Every chat is saved, so you can come back to it.

   • Resume reopens an old chat with its full context.
   • Fork makes a copy, so you can try a different approach without losing the original.
   • Archive tidies away finished work.

   In the app you can rename, pin and search your chats.

   Think of it like…

   Resume is reopening a saved document. Fork is "Save As" a copy before experimenting.

   Try this

   /fork

   Copy

   You’re done with this step when

   You know how to get back to an old chat, and how to try an idea without losing the original.

   For the curious

   In the terminal:

   • codex resume opens a picker, and codex resume --last reopens your newest chat.
   • /fork, or Esc twice on an empty message box, forks from an earlier message.
   • /archive hides a chat. /delete removes it permanently.

   I did this

4. Step 8 of 16

   AGENTS.md: the rules it reads before any work

   Each chat starts fresh, so anything Codex should always know belongs in a plain text file called AGENTS.md. Codex reads it before doing any work in that folder.

   Good things to put in it:

   • what the project is
   • how to check the work
   • your preferences
   • things it must never do, like "ask before deleting files"

   Keep it short and accurate. OpenAI's own advice: a short, accurate AGENTS.md beats a long one.

   The easiest start is /init, which writes a draft for you to edit.

   Think of it like…

   Company handbook, then department handbook, then the sticky note on your own desk. The most local one wins.

   See an example: An AGENTS.md for your practice folder ▸See an example: How AGENTS.md files stack up ▸

   Try this

   /init

   Copy

   You’re done with this step when

   Your practice folder has an AGENTS.md, and you added one rule of your own to it.

   For the curious

   Files layer from general to specific:

   • ~/.codex/AGENTS.md applies everywhere.
   • The project's own file comes next.
   • A file in a subfolder adds to it, and the closest file wins.

   An AGENTS.override.md replaces the normal file at its level. The combined size defaults to 32 KiB. Codex re-reads the files each session, so restart to pick up changes.

   Claude Code also reads AGENTS.md when a project has no CLAUDE.md.

   I did this

5. Step 9 of 16

   Memory: helpful recall, not rules

   ChatGPT can remember things about you across chats, and Codex on your computer has its own separate memory, which is off until you turn it on in Settings.

   OpenAI is clear about the difference:

   • Memory is a helpful recall layer. Use it for preferences Codex picks up as you work.
   • AGENTS.md is where rules that must always apply belong.

   A ChatGPT project is different again: a folder of related chats, files and instructions. It does not give access to a folder on your computer.

   Think of it like…

   AGENTS.md is the laminated rules sheet on the workbench. Memory is a colleague's recollection: usually right, not to be relied on for the rules.

   You’re done with this step when

   You can say which belongs in AGENTS.md and which you would leave to memory.

   I did this

6. Step 10 of 16

   Sandbox and approvals: what it may touch

   Two controls work together:

   • The sandbox decides what Codex can technically reach.
   • Approvals decide when it must stop and ask before going further.

   Internet access is off by default. The modes:

   • Ask for approval (recommended) works inside your folder and asks before going beyond it.
   • Approve for me passes those requests to an automatic reviewer instead of you.
   • Full access has no limits: your whole computer and the internet. Avoid it.

   Think of it like…

   The sandbox is a fenced yard. Approvals are the gate, where it rings the bell.

   Try this

   /permissions

   Copy

   You’re done with this step when

   Your permissions are set to Ask for approval, and you know what each mode allows.

   For the curious

   In config these are:

   • sandbox_mode: read-only, workspace-write or danger-full-access
   • approval_policy: on-request or never

   Codex enforces the sandbox with the operating system's own isolation. Experimental rules in ~/.codex/rules/ can allow, prompt for or forbid specific commands, and the strictest rule wins.

   I did this

1. Step 11 of 16

   Plan first, and set a goal

   For anything bigger than one file, do not let it dive in.

   • /plan switches to plan mode: Codex looks around, asks questions and proposes steps before touching anything.
   • /goal sets the outcome and what "done" means, so a long task keeps working towards it.

   See an example: A goal with a clear "done" ▸

   Try this

   /plan I want to add a contact form to the bakery page. Propose the steps and wait for my OK.

   Copy

   You’re done with this step when

   Codex showed you a plan and waited for your go-ahead.

   I did this

2. Step 12 of 16

   Skills: teach it a job once

   A skill is a folder with a SKILL.md file: written instructions for one job, such as "how we write the weekly status update".

   Codex sees only each skill's name and short description until a task matches, then loads the full instructions. So skills cost almost nothing until they are needed.

   • Call a skill by name with $, like $weekly-status.
   • Or let Codex pick it when your request fits.
   • Codex comes with $skill-creator to help you write new ones.

   Think of it like…

   A recipe box. You read the labels and only pull out the full card when you are cooking that dish.

   See an example: A skill file: weekly status ▸

   Try this

   $skill-creator Make a skill called weekly-status that turns my rough notes into: done, next, blocked.

   Copy

   You’re done with this step when

   You made a skill and used it by name.

   For the curious

   Skills live in .agents/skills inside a project, or ~/.agents/skills for you personally. OpenAI also ships some built in.

   Write the description with the main use and trigger words first. Skills are shared as plugins.

   I did this

3. Step 13 of 16

   Subagents and worktrees: helpers in parallel

   A subagent is a helper agent Codex sends off with one job. It works with its own context and reports back a summary, so your main chat stays clean.

   You can ask for several at once. In Codex on your computer, they start only when you, your AGENTS.md or a skill asks for them.

   To stop parallel helpers tripping over each other, Codex can work in a worktree: a separate copy of your project, so each one edits its own files.

   Think of it like…

   A project lead sending three researchers out with their own copies of the files, then reading their one-page reports.

   Try this

   Review this folder with parallel subagents: one for spelling and tone, one for broken links, one for layout on phones. Summarise by category.

   Copy

   You’re done with this step when

   You can explain why a subagent keeps your main chat clean.

   For the curious

   Built-in agent types are default, worker and explorer. Custom agents are TOML files in .codex/agents/, and /agent switches between agent threads.

   Subagents use more of your allowance than a single agent.

   I did this

4. Step 14 of 16

   Apps, plugins and MCP: plugging into your tools

   On its own, Codex sees your folder. Apps (connectors) and MCP let it reach other tools, such as your docs, issue tracker or design files.

   MCP, the Model Context Protocol, is an open standard for plugging tools into AI agents. Plugins are installable bundles of skills, MCP connections and more, and ChatGPT and Codex share one plugin directory.

   Connect only what a task needs. A connection acts with your access.

   Think of it like…

   MCP is a universal power socket. A plugin is an appliance that arrives with its own plug and instruction booklet.

   See an example: Adding an MCP server ▸

   You’re done with this step when

   You can explain what MCP is for, and where you would add a connection.

   For the curious

   In the terminal:

   • codex mcp add <name> -- <command> adds a server.
   • /mcp lists your servers.
   • /apps inserts an app.

   Servers are set in ~/.codex/config.toml. You can limit which of a server's tools are allowed and when they need approval.

   I did this

5. Step 15 of 16

   Models, effort and review

   Models, switched with /model:

   • GPT-6 Astra is the most capable.
   • GPT-6.1 Sol is the recommended everyday default.
   • GPT-6 Luna is the most efficient, and the one on the free plans.

   Reasoning effort is how long it thinks before answering. More effort helps on hard problems but is slower and uses more of your allowance.

   /review starts a separate reviewer that reports problems without changing your files. It is a second pair of eyes before you keep anything.

   Think of it like…

   Effort is how long a consultant mulls before replying, and you pay for the mulling time.

   Try this

   /review

   Copy

   You’re done with this step when

   You know which model you are using, and you ran a review on your changes.

   For the curious

   Usage is shared with Work and resets every five hours, with weekly limits too. On Plus, Luna allows far more messages than Astra. /status and /usage show where you stand.

   On GitHub, commenting @codex review on a pull request asks Codex for a review.

   I did this

6. Step 16 of 16

   Put it together on a real job

   Pick one small, real task: clean up a messy spreadsheet export, write a plain-English guide to a project a colleague left you, or explain a folder of files.

   Use what you learned:

   1. Work on a copy.
   2. Add an AGENTS.md with your rules.
   3. Plan first.
   4. Review every change.
   5. Start a new chat when you move on.

   Try this

   Clean this CSV: remove duplicate rows, fix the date formats and trim extra spaces. Save a cleaned copy and do not change the original.

   Copy

   You’re done with this step when

   Codex finished a job you needed, and you reviewed the result before keeping it.

   I did this

Worth knowing

• Codex usage is shared with ChatGPT Work and has limits that reset. Larger models and subagents use it faster.
• Never paste passwords or API keys into a prompt or a project file.
• Cloud tasks run on OpenAI’s servers and cannot see your local files or the sites you are logged in to.

Where this comes from

• Using ChatGPT: Chat, Work and Codex
• Codex quickstart
• Pricing and plans
• AGENTS.md
• Memories
• Glossary (context window)
• Slash commands
• Permission modes
• Skills
• Subagents
• MCP
• Models
• Code review

Checked against the sources on 2026-10-08. These tools change quickly.