What Is ChatGPT Work? A Plain English Guide for Business Teams

ChatGPT Work is an agent that lives in the ChatGPT desktop app, reads the files on your computer, connects to the tools you already use, and hands back a finished piece of work. While "regular" ChatGPT gives you a response and you go and do the job, ChatGPT Work does the job for you.

That is the whole distinction, and it changes what you should be using ChatGPT for. This guide covers what it actually is, why you should use the desktop app to get the most out of the product, how projects and files work, and the two concepts—plugins and skills—that enable people to work most effectively with it.

How do Chat, Work and Codex fit together?

Let's start with what you probably already know. Regular ChatGPT is an AI chat, and its role is to be a thinking partner. Questions, explanations, brainstorming, a quick second opinion. You give it a prompt, you get an answer, and in most cases you then go and do the work.

ChatGPT Work is different in kind. It's an agent, which means you delegate a task rather than ask a question. It takes the task, makes a plan for itself, and works through the steps until the job is done.

This is where the market is going rather than a niche feature. McKinsey's 2025 State of AI survey found 62% of organisations at least experimenting with AI agents, though fewer than 10% have scaled them in any one function.

Under the hood, both run the same agent. Ask ChatGPT Work what system it's running and you'll see both are the same Codex agent, living in two product interfaces focused on different areas.

Work and Codex are the same engine pointed at different jobs. Codex is aimed at writing software. Work is aimed at knowledge work: multi-step tasks across your tools, pulling data for analysis, updating a CRM, producing spreadsheets and documents. If you sit in finance, legal, HR, marketing, sales or operations, Work is the one you want.

Claude draws the same line between Chat, Cowork and Code, for the same reason. The practical test for whether something belongs in Work rather than Chat: can you describe the deliverable? "What should go in a cohort analysis" is a chat. "Build a cohort analysis from these five exports, with every cell as a traceable formula" is a job.

The desktop app is not optional

ChatGPT Work runs in two places, and they are not equivalent.

The product page states it plainly: Work is available to all plans on the desktop app, and to Plus, Pro, Business, Enterprise and Edu on web and mobile. So even a free account gets it, as long as you download the desktop app.

Even if your plan allows the web version, use the desktop app anyway. The web version has the Chat and Work tabs, but Work there does not carry the full feature set, and it will prompt you to open the desktop app regardless.

The reason is that Work has two modes and only the desktop app gives you both.

Local mode runs on your computer. It reads your files, and it gives you the built-in browser inside the desktop app.

Cloud mode runs on OpenAI's servers. Its one real advantage is that a task keeps running with your machine off. That is roughly where the advantages stop, because it cannot touch your local files.

I use local mode almost exclusively, and the file access is why. On a set of five CSV exports running to hundreds of thousands of rows, the browser version refuses the upload outright. The desktop app reads the same files straight off the disk without uploading them anywhere.

Projects are how you hand over your files

The habit to build: every task you give ChatGPT Work should live in a project. This is a different discipline from regular ChatGPT, where a chat is just a chat.

When you start a task you can choose a project. Create one, name it, and select the source folders ChatGPT can read and edit.

There are two ways to give it files, and they are not the same thing. You can attach files and folders with the plus button, which works but gives you only the built-in browser in the right-hand panel. Set the same thing up as a project instead and that panel gains a Files tab showing everything in your project folders. That matters more than it sounds, because Work is usually creating files as well as reading them, and you want them contained somewhere you can watch.

A project can hold more than one source folder, with one constraint worth remembering: it can only have one primary folder. That is where it looks for the AGENTS.md file covered below, and it matters for skills too.

Here is a real run. I pointed a project called "accounting" at a folder of receipt images and briefed it: extract the details from every image, build an Excel file, then create a "processed receipts" subfolder and sort each image into subfolders by the month the spend happened.

It made its own plan and showed it: inventory the images, visually inspect each receipt, build and verify the Excel, move the validated files, then report what was logged, what needed review and what was archived. Claude produces the same class of output through its own file creation environment. After a little over six minutes it was done. One row per expense, with a confidence score I had asked for, and a single receipt flagged for review because the printed VAT rate was too faint to read confidently.

That last detail is the one to notice. It did not guess. It flagged.

Because these runs are long, watch the plan when it appears. If you disagree with it, stop it and correct it rather than waiting to be disappointed.

AGENTS.md stops you repeating yourself

AGENTS.md is a plain markdown file in your project's primary folder holding persistent instructions. Whenever you start a new session in that project, the first thing Work does is read it.

The effect is easy to measure. In a project without one, I asked where our Notion database for YouTube videos lives. It took 43 seconds, with several calls to Notion and multiple pages searched before it landed on the right one. The same question in a project with an AGENTS.md came back in two or three seconds and searched nothing at all, because the file already held the link.

Those 43 seconds are usage you paid for and did not need to spend. Multiply it across a team and it stops being trivial.

The rule of thumb: if you find yourself typing the same instruction twice, it belongs in AGENTS.md. There is a built-in /init command, but I find it overkill and tuned towards Codex, so you end up with far more than you need. Asking for what you actually want works better: create an AGENTS.md noting that it should never make a new Excel file, but add new receipts to the existing one and check for duplicates.

Plugins: apps tell it where, skills tell it how

This is the concept that takes you furthest, and it is also where most of the confusion sits.

In the left-hand menu you will find Plugins, described as working with ChatGPT across your favourite tools. That phrasing leads people to assume a plugin is just a connector. It is not.

A plugin is an installable package containing two things. Apps, which connect ChatGPT to a tool, much like Claude's connectors. And skills, which teach it repeatable workflows. An app tells it where to work. A skill teaches it how to work.

That is why the plugin list mixes tools like Gmail and Google Drive with things like Presentations. Presentations is not a tool and connects nothing; it is a single skill that produces polished decks. The Gmail plugin has both: the app connects your account, the skill teaches it how to search your mail and write messages.

Installing has a trap. Install Gmail and it immediately asks you to connect your account, because the plugin holds one app. Install Data Analytics and nothing pops up at all, which makes it look finished. It is not. Click into the plugin and you will find it suggests more than 23 apps to connect, because data analysis usually means pulling from your existing tools. Go through the list and connect what you actually use.

Most plugins also let you toggle individual skills on and off. Turn off the ones you will never use, so a stray keyword cannot trigger them.

To direct it rather than letting it choose: @ pulls up plugins, / pulls up skills, and $ lists skills only.

When your tool has no plugin

You are not stuck. The mechanism is MCP, the Model Context Protocol, an open standard for connecting AI applications to external systems.

Search your tool's name plus "MCP", find the server URL in its documentation, then add it through the web version under Settings, Plugins, Browse plugins. If you cannot see a plus button there, enable developer mode in the same settings area first. Depending on your plan you may need an admin to do it.

Two things caught me out. The connector has to be published from the three-dot menu before it appears in the desktop app, which is well hidden. And the plus button does not exist in the desktop app at all, so this is a web-version job.

Skills are where the compounding happens

Skills let you save a workflow and run it on demand, the same way every time. Anthropic's documentation on how skills load describes the same design on the Claude side: metadata always in context, instructions only when triggered, bundled scripts run without their code entering the context window.

The difference is measurable. I asked Work to build a marketing email in Customer.io. With no skill attached, it took over 8 minutes and produced something functional but generic, nothing like our template. The identical prompt with our email skill enabled finished in 1 minute 51 seconds and came back in exact 9x style: logo top left, our cream background, our button style, the right footer links.

You do not write these by hand. The process that works:

1. Do the task with ChatGPT Work step by step, once.

2. Refine until you are happy. When it gets something wrong, tell it rather than quietly fixing it yourself.

3. Have it save the result as a skill, using the built-in /skill creator.

4. Test it and keep correcting.

Doing it in that order matters. Starting from the skill creator with a blank page risks building careful instructions for something the tool cannot actually do.

Skills can also call other skills. When I saved our blog image workflow, it noted for itself that it should run the built-in image generation skill whenever it creates one. Build the smallest reliable piece, then compose.

What ChatGPT Work costs

Work is on every plan on the desktop app, including the free tier, with paid plans adding web and mobile access and much higher limits. Free and Go get limited access.

A few features are not on free at all, scheduled tasks and Sites among them, so check before you design a workflow around one. File creation draws on the same usage allowance as normal chats, and it draws harder.

Common questions

What is ChatGPT Work in one sentence? An agent in the ChatGPT desktop app that takes a whole task, works on your real files and tools, and returns a finished deliverable.

Is ChatGPT Work the same as Codex? Same agent underneath, different product surface. Codex is pointed at writing software; Work is pointed at knowledge work.

Do I need a paid plan? No. Work is available on the free plan through the desktop app, with usage limits. Paid plans add web and mobile and considerably more headroom.

Can it work on files on my computer? Yes, in local mode on the desktop app, and only in the folders you connect to a project. Cloud mode cannot reach local files. Claude Cowork takes a stricter default here, staying folder-scoped for reading as well as writing.

How is it different from a custom GPT? A custom GPT holds a long instruction set in context at all times. Work loads a skill only when the task calls for it, and it can act on your files and tools rather than only producing text.

Why is my task taking so long? Because it is doing the work rather than answering. Runs of six minutes to over an hour are normal. Watch the plan when it appears and stop it early if the approach is wrong.