How to connect ChatGPT to your business tools with MCP
To connect ChatGPT to a business tool, add the tool's server address and sign in through your browser. ChatGPT can then read the records that product exposes and, when permitted, change them. It sees only the connected account. Start with questions, check the answers, and try actions only when you trust the boundary.
Key takeaways
- •Connecting is a URL and a browser sign-in. There is no file to upload and no key to paste into a chat box.
- •The assistant reaches one product, in one account, with your permissions. Everything you did not connect stays invisible to it.
- •ChatGPT has moved this setting more than once, and its own help pages have lagged behind the app.
- •For multi-step work, Codex in Code mode is the stronger half of the pair. Plain ChatGPT is at its best asking questions of an account.
- •Test with questions before you test with actions. A wrong answer costs nothing.
What connecting means
An MCP server is an adapter a product publishes so that an AI assistant can use that product's own features directly. Your side of it is small. You paste a URL, a browser window opens on the vendor's domain, you approve what is being granted, and the assistant now has a menu of tools it can call in your account.
Three things are worth stating plainly, because the usual expectations are the wrong ones. Nothing is copied into the model. The assistant does not learn your business between sessions; it looks things up at the moment you ask, the way you would. And the results arrive in the conversation, so whatever your AI client does with a chat transcript, it now also does with the rows those tools return.
The protocol side of this is covered in what an MCP server is. This article is about the ChatGPT half of it.
The limits, first
Most people arriving at this question want ChatGPT to answer questions about their own business. It can. Six things get in the way, and it is better to meet them here than an hour into a setup.
-
One product, one account
A connection reaches the product it belongs to, in the account you approved. Your other tools, your inbox and your files are not part of it.
-
The setting moves
ChatGPT has relocated its MCP section more than once, and its help pages have lagged behind the app. Trust the app, not a screenshot.
-
Codex needs a helper
Codex has no native HTTP transport. It reaches a remote server through
mcp-remote, which runs on your own machine. -
Chat is not a build tool
Our own setup guide recommends plain ChatGPT for asking questions of an account, and Codex in Code mode for multi-step work.
-
Three calls at a time
The AXL server takes three requests in flight per caller and refuses the rest. A large job is paced rather than instant.
-
Money and rights stay yours
Billing writes, API-token minting and permission changes are off the agent's surface. It cannot raise a spend limit or promote itself.
The last card is a design decision rather than an obstacle. Twenty-one sensitive operations are removed from the AI surface in the server, not discouraged in a prompt, which is the difference between a rule and a suggestion. The reasoning is written out in what an agent can never do.
What it can reach once connected
On the AXL side the connection is not a narrow read-only window. The server publishes the admin surface of your account: contacts and deals, courses and lessons, site pages, forms, email, automations, orders and payments.
Coverage is only half of it. A schema tells a model which fields exist; it does not say that updating a contact's tags replaces the whole set, or that a bulk action with an empty filter means every record in the account. Those are the notes attached to the recipes, and they are the difference between a call that succeeds and a call that does the right thing.
Two things stay deliberate. Authoring and publishing are separate calls, so an agent can build a page without putting it live and prepare a campaign without sending it. And the agent inherits the permissions of the person who signed in, inside the one account the consent screen named.
If you want the wider picture rather than the ChatGPT-shaped slice of it, see what an AI agent can run end to end on AXL: pages, products, email and automation in one account.
Start with questions, not actions
The first hour should cost nothing if it goes wrong. Ask things you can check yourself in a screen you already know.
- •“How many contacts were created last week, and where did they come from?”
- •“Which of my automations have never run?”
- •“List the orders that were paid but never got course access.”
- •“What did we earn last month, split by product and by channel?”
- •“Which students paid twice and should be refunded?”
“How many contacts were created last week, and where did they come from?”
The agent reads the menu, finds the operation, then calls it in your account.
Check the first few answers against the product's own screens. Then move to actions one at a time, and ask for a summary of what changed before you ask for the next thing.
When ChatGPT is the wrong client
Two jobs, two clients. Asking questions of an account works well in the ChatGPT window. Building something across several steps works better in Codex Code mode, or in Claude, Cursor, Windsurf, VS Code or Cline, which all speak the same protocol to the same server.
The server does not care which one connects, and adding a second client does not disconnect the first. So if ChatGPT is where you already work, start there, and add another client the day a job outgrows the chat window.
The steps for both halves of the OpenAI pair are on the ChatGPT and Codex setup page, which is kept in step with the screen inside the product. The Claude path is on its own page.
Is it worth the setup?
The setup is a few minutes. The honest test is how often you ask questions that cross two screens.
If your reporting question is always the same one, a saved report answers it faster and costs nothing. If you ask a different question every week, and the answer currently means exporting two lists and joining them by hand in a spreadsheet, that is the work this removes.
The second case is setup rather than analysis. Launching an offer touches a page, a product, a payment step, a confirmation email and a follow-up sequence. An agent can assemble those pieces while you keep the decision to publish. That is a different reason to connect, and it is the one that usually justifies the afternoon.
FAQ
Does this send my business data to my AI provider?
The records a tool returns arrive in the conversation, the same as anything you paste in yourself. So whatever your AI client's settings say about conversation data, they now cover those records too. That is a question for your client's data controls, not for the server you connected.
Do I need a developer?
For ChatGPT, no. It is a URL and a browser sign-in. The Codex route runs mcp-remote on your own machine, so that one wants someone comfortable with a config file.
Can the agent break something?
It acts as you, with your permissions, so it can do what you can do. Sensitive operations are removed from its surface, publishing and sending stay separate steps you ask for, and every change lands in the product where you can see it. Start read-only anyway.
Which account does it get access to?
The one named on the consent screen. If you administer more than one, switch there before you approve, because the token is scoped to that account.
There is no MCP section in my ChatGPT settings. Now what?
Look for connectors or developer settings in your build, since the section has moved before and may move again. If the path on our setup page no longer matches what you see, tell us and we will fix the page.
Ask ChatGPT something only your account knows.
One URL and a browser sign-in, and the assistant you already use can read your contacts, courses, orders and campaigns. Publishing and sending stay separate, deliberate steps.
- •API keys: the account-side permissions model an agent inherits.
- •Team and departments: how rights are granted, which is what the agent gets and no more.