What is Codex and how to use it for your business

You’ve seen Codex appear alongside ChatGPT. You’ve heard people talking about AI agents. Now you want to know what any of it can actually do for your business.

Codex is OpenAI’s coding agent. It can work with files, write and change code, run commands and check its work. For a business owner, that opens up practical possibilities: building a small tool, improving a website or creating a workflow around a recurring job.

The useful starting point is a job you can explain and a result you can check.

Think of it as briefing a new team member. You’d explain your business, show them the relevant material and agree on what finished looks like. AI needs that context too.

What is Codex?

Codex is an AI agent designed for software development. You describe a goal, and it can use the files and tools available to it to work towards that goal.

An agent can take several steps: inspect a project, propose an approach, make changes and run checks. What it can access and do depends on the environment, tools and permissions you give it.

OpenAI’s current quickstart separates Codex for software development from ChatGPT Work for research, analysis and finished documents. That distinction helps you choose where to begin.

If you want a report, presentation or article, ChatGPT Work is the natural starting point. If you want to build a calculator, change a website’s code or develop a custom business tool, Codex is the more relevant choice.

Codex can also handle supporting work with files and connected tools. In the article example below, it’s part of a wider business workflow. That doesn’t mean every business task needs Codex.

Is Codex free?

There’s a limited free entry point. As checked on 1 October 2026, OpenAI lists Codex access with GPT-6 Luna in the desktop app on the Free plan, subject to rollout.

Paid plans offer additional access and allowances. ChatGPT Work and Codex share usage, and the amount a task consumes depends on the model, complexity and tools involved. Check OpenAI’s current plan information and your account before choosing a plan.

Free access is a way to explore. It isn’t a promise of unlimited work or access to every feature.

How do Codex, ChatGPT Work and Dot fit together?

They serve different purposes within OpenAI’s tools.

• ChatGPT Chat is where you ask questions and think through ideas in conversation.
• ChatGPT Work carries research, analysis and file creation through to a result you can review.
• Codex focuses on software development and work involving code.
• Your Dot is an always-on agent that can keep responsibilities moving between conversations and assign work to other tasks.

OpenAI says Dot can research, prepare documents and start Work or Codex tasks. It has a cloud computer; using your local computer requires a separate connection and that computer being available. Dots are gradually rolling out to eligible accounts. See OpenAI’s Dots guide.

Think of Dot as the coordinator. You can give it an ongoing responsibility, then use the appropriate worker for each job. You still need clear instructions, suitable access and a review process.

How to start using Codex for your business

Choose one useful job first. Then choose where it should happen.

For local work, OpenAI’s desktop setup guide explains how to sign in, select a project or folder and select Codex for software development. Only give it access to the files relevant to that work.

You might begin with a small website calculator, a simple internal tool or a script that organises a recurring export. These are possible starting projects, not claims that any tool can finish them without testing.

For the walkthrough below, we’ll use an article workflow to make the principles easy to follow: one job, current business knowledge, relevant tools and a result you can review. Those same principles apply when you’re building software.

Give your agent one clear job

Start with an outcome, then describe what finished looks like.

For an article task, that could mean: research a named topic, save an evidence-backed brief and bring it back for your approval. Writing the article is the next stage. Publishing it is another decision again.

Those stages matter. A saved research brief is a useful finished result. It isn’t an approved article, and it isn’t a published page.

My recommendation is to make the first job clear enough to assess before you add more jobs and connections. A small, useful result gives you something concrete to review.

Write down the output you want, where it should be saved and what the agent must bring back to you. Give it a stopping point too: ask when essential information is missing, and return for approval before the next stage.

Here’s a starter brief you can adapt:

“Use the approved business and brand files in this project. Research [topic] for [reader], using current primary sources. Save a brief with the key questions, source links and any claims you couldn’t verify. Show me the proposed article angle before writing. Keep private business information out of the output. Don’t publish or send anything.”

Replace the bracketed details and name the actual files you want it to use. If you’re building a tool, specify what it should do, who’ll use it and how you’ll test that it works.

Give it your business knowledge

Your agent needs to understand the business behind the task.

I use an AI Brain for Business to organise that intelligence. Business DNA captures the strategy, audience, offers and priorities. Brand DNA captures the voice and how the business should communicate. Relevant knowledge files add the detail needed for the particular job.

The AI Workspace gives those files a clear home. It helps you find the current sources and file the work that comes back.

For an article, the useful context might include your intended reader, your approved offer, your voice guidance and the research brief. For a different task, the relevant files will be different.

My method is EXTRACT, ORGANISE, AUTOMATE. Capture the business knowledge, organise it so it can be used, then build the work around it.

Keep the context current. If two files disagree about an offer or a price, tell the agent which source takes priority. A connection to a folder doesn’t make every file accurate, and storing files doesn’t create unlimited or permanent model memory.

Add the skills and apps the job needs

A skill gives the agent a reusable way to handle a particular kind of work.

OpenAI’s skills documentation explains that a skill contains instructions in a SKILL.md file and can include references, scripts or assets. Skills can be selected explicitly or used when the task matches their description.

For this article workflow, the Researcher gathers evidence. The Cornerstone Article Writer uses an approved brief to create the article. A publishing workflow handles the approved copy and destination later.

Connected apps provide access to other systems when the connection and permissions allow it. A research tool might supply keyword estimates. A website connection might let you prepare or publish a blog post.

Connect what the job needs. Read the tool’s capabilities before assuming it can do the whole workflow. Being able to read a website’s blog records doesn’t prove the connection can publish a new post successfully.

Set the review boundary before you run it

Tell the agent what it can do independently and what needs your approval.

Local research and drafting are different from sending an email, changing a customer record or publishing on your website. Name those boundaries in the task.

For example, your article instruction can allow the agent to read the relevant business files, research public sources and save a draft locally. It can also require it to show you the exact copy and destination before creating anything on your live website.

Keep sensitive material out of public outputs. Business knowledge can guide the work without being copied into the article.

You remain responsible for the decisions. The agent should make it easier to review the work, with clear sources, visible uncertainties and a saved version you can open.

Run, review and verify the result

Open the result. Check whether it does the job you asked for.

For an article, check the reader, the claims, the source links, the voice and the next action. Ask for a specific edit, then check the saved version again.

Research needs review too. Keyword-tool estimates can help you understand search interest, but they aren’t exact counts of your audience. They don’t tell you what people privately ask ChatGPT, Claude or Perplexity.

When you approve a later external action, check that it actually happened. For a published article, that means opening the public page and checking the content. A tool saying it received a request is only part of the evidence.

Useful work has a visible result: the brief is saved, the article is reviewed or the public page is verified. Keep those states clear.

Build the foundation, then expand the work

Once one job works, you have something concrete to improve. You can refine the instructions, update the business knowledge or add the next relevant capability.

That’s the work we do inside Content Catalyst: build your AI Brain and AI Workspace, then use skills, workflows and supervised agents around practical business priorities.

You have live guidance and a library of tools to help you implement the work, while keeping your judgement and approval in the process.

If you want help setting up AI around your strategy, admin and marketing, explore Content Catalyst.