How to actually use Codex to run your business (not just write code)

Most people who install Codex for the first time treat it like a fancy autocomplete tool. They ask it to fix a bug, it fixes the bug, they close the terminal and go back to whatever they were doing before. That's the equivalent of buying a forklift and using it to carry one box at a time.

Codex isn't really a coding tool anymore — it's an agent that happens to be really good at code. And once you stop thinking of it as "the thing that writes my Python scripts" and start thinking of it as "the employee who never sleeps and does exactly what I tell it to," the whole calculus of what's worth automating in your business changes.

This guide walks through what Codex actually is in 2026, how to set it up properly, and how to use it for things that have nothing to do with shipping software — from client-facing content to internal tooling that used to require hiring a developer.

What is Codex, really?

The official Codex documentation lives at developers.openai.com/codex and it's worth bookmarking, because the tool changes fast enough that anything older than a few months is already stale. At its core, Codex CLI is described in its repository as a "lightweight coding agent that runs in your terminal" that reads your repo, runs commands, and edits files in place while you watch.

But the CLI is just one surface. Codex runs in four main places, and the right one depends on how much you want to watch the agent work versus hand it a task and walk away. There's the terminal, a desktop app, an IDE extension for VS Code and JetBrains, and a cloud version that works in a sandboxed environment and hands you back a proposed change instead of touching your local files. The surfaces share one account, so you can start a task in the editor and hand it off to the cloud to finish — the cloud agent works on its own in a remote environment and hands back a proposed change rather than editing your local files.

That last point matters more than it sounds. It means you can kick off a task from your phone, walk away, and come back to a finished pull request. This is the difference between "AI helps me code faster" and "AI does the thing while I'm at lunch."

How do you actually install and set it up?

Getting Codex running takes about five minutes if you've got the prerequisites sorted. According to setup documentation, you'll need a ChatGPT account with Codex access — included with paid ChatGPT plans, or use an OpenAI API key — plus Node.js 18+ for npm install, or Homebrew on macOS.

Here's the basic flow:

1. Install the CLI. Use npm or Homebrew depending on your platform. The official openai/codex repository documents Homebrew installation, and the CLI can also be installed with npm and then started with the codex command.
2. Authenticate. The first time you run codex, you sign in with either your ChatGPT account or an API key — the paid ChatGPT plans include Codex access, so most people use the account login.
3. Launch it inside a project. From a project's root, run codex, and it launches an interactive terminal interface, inspects your repository, and works within the directory you point it at.
4. Pick your approval mode. This is the part people skip and later regret. Before Codex changes files or runs commands, you choose the approval mode that matches your comfort level. Codex CLI documentation breaks this into three tiers: Auto, the default, lets Codex read, edit, and run commands within your working directory but asks before anything outside that scope; Read-only keeps Codex consultative; and Full Access grants broader autonomy including network access.

If you're new to this, start restrictive. Start in a mode that asks before acting, and loosen it only once you trust the pattern. You can switch modes mid-session with /permissions whenever your comfort level changes.

Why does the AGENTS.md file matter so much?

This is the single biggest lever most beginners never pull. If you only remember one thing from this article, make it this one.

AGENTS.md is a plain markdown file you drop into the root of a project that tells Codex — and pretty much every other coding agent now — how your project actually works. The format is deliberately minimal: plain Markdown with no required schema, no YAML front matter, and no special syntax. Where a README.md explains a project to humans, AGENTS.md explains it to agents.

What goes in it? Commit messages or pull request guidelines, security gotchas, large datasets, deployment steps — anything you'd tell a new teammate belongs here too. Codex reads it automatically the moment a session starts. Codex reads it automatically at session start, before any work begins.

The adoption numbers back up how central this has become. AGENTS.md was released by OpenAI in August 2025 and transferred to the Linux Foundation's Agentic AI Foundation alongside Anthropic's MCP and Block's Goose, and by May 2026 it had been adopted by more than 60,000 open-source repositories. It's not a Codex-only thing anymore either — AGENTS.md emerged from collaborative efforts across the AI software development ecosystem, including OpenAI Codex, Amp, Jules from Google, Cursor, and Factory.

One underrated feature: nested files. You can place another AGENTS.md inside each package, and agents automatically read the nearest file in the directory tree, so the closest one takes precedence and every subproject can ship tailored instructions. If you're running multiple client projects out of one repo structure, this means each one can have its own rules without you re-explaining context every session.

Worth noting: this file reduces mistakes, it doesn't eliminate them. AGENTS.md reduces rule violations by making rules prominent and always-on, but it doesn't eliminate them — for actions where rule violations would be irreversible or high-cost, the right safeguard is a combination of AGENTS.md instructions and explicit approval gates. Don't skip human review on anything that touches production or client data.

Which model should you actually use?

OpenAI ships multiple models under the Codex umbrella, and picking the wrong one is a common way to waste money or get sloppy output. The documentation recommends starting with gpt-5.4 for most workflows. You can switch on the fly — either at launch with a flag or mid-session using the /model slash command inside the terminal interface.

The general rule: don't use the smaller, faster model for anything high-stakes. Use it for quick lookups, simple edits, and read-only exploration. Reserve the heavier reasoning model for anything that touches money, client-facing output, or infrastructure you can't easily roll back.

What can Codex actually do outside of coding?

This is where things get interesting for non-developers. Because Codex can read files, browse the web, and interact with connected tools, it stops being "a thing that writes code" and becomes "a thing that produces finished work." As one guide to the desktop app puts it: you describe the result you want in plain English, Codex makes a plan, works across your files, web, and connected tools, and hands you back a finished document, deck, spreadsheet, or even a working app — and the best part is you don't need to be a developer.

In practice, this opens up a handful of genuinely useful business workflows:

• Internal tools without hiring a developer. A calendar assistant, a lead-scoring script, an invoice generator — things that used to require a freelancer now take an afternoon of back-and-forth with Codex.
• Automated reporting. Point it at your data exports and have it build a script that formats a weekly report, no manual spreadsheet wrangling required.
• Version-controlled experiments. Once something works, ask Codex to push it to a GitHub repository so you're not relying on a file sitting on one laptop.
• Content and copy in your own voice. With a well-written AGENTS.md describing your tone, structure preferences, and dos/don'ts, Codex-driven writing workflows stop sounding like generic AI output and start sounding like your brand.

What should beginners watch out for?

A few practical warnings worth knowing before you hand Codex the keys:

Don't skip the review step. OpenAI's own Codex introduction states that it still remains essential for users to manually review and validate all agent-generated code before integration and execution. This applies just as much to a client email draft as it does to a database migration.

Windows support is still catching up. macOS or Linux is recommended, since Windows support is experimental — for the best Windows experience, use Codex in a WSL workspace.

Skills are different from AGENTS.md. If you find yourself repeating the same instructions over and over, that's a sign you need a Skill, not another paragraph in your instructions file. When you find yourself writing the same prompt pattern repeatedly, extract it into a skill — a directory containing a SKILL.md file with YAML frontmatter, plus optional scripts, references, and templates. Think of AGENTS.md as the constitution and Skills as the reusable playbooks.

How do you get started this week?

Don't try to automate your entire business on day one. Pick one recurring task that currently eats an hour of your week — a report, a repetitive piece of internal software, a content workflow — and build a single AGENTS.md file around it. Get comfortable with approval modes before you loosen them. Then, once that one workflow runs clean for a few weeks, expand.

The official Codex documentation and the AGENTS.md specification are both free, both actively maintained, and both worth reading in full before you start pointing an agent at anything that matters. The people getting the most out of Codex right now aren't the best programmers — they're the ones who wrote the clearest instructions.