How to build a ChatGPT app: Apps SDK, approval and cost in 2026

To build a ChatGPT app in 2026 you use OpenAI's Apps SDK, and the part people underestimate is how little of it is new code. A ChatGPT app is an MCP server over your existing API plus a few small interface components that render inside the chat; you test it in developer mode, then submit it to the ChatGPT app directory for OpenAI's review. A focused first version takes 1 to 2 weeks with an AI-first team, 3 times cheaper and 3 times faster than a traditional hand-written build.

The reason it is worth doing now: OpenAI opened app submissions to all developers on December 17, 2025, and the directory at chatgpt.com/apps sits in the sidebar of a product that already had 800 million weekly users when apps launched. Booking.com, Canva, Figma, Spotify and Zillow were in the pilot; dozens of brands followed. For most categories, the shelf is still mostly empty.

What a ChatGPT app actually is

Strip away the launch video and an app inside ChatGPT has two layers:

• Tools. An MCP server lists the actions ChatGPT may call, such as search_listings or track_parcel, each described in plain language so the model knows when to use it. This is the same Model Context Protocol that Claude, Gemini and Copilot use, so the backend work carries over. Our explainer what is an MCP server covers the protocol and its security risks in depth.
• Interface. Optional components (a product card, a carousel, a map, a form) that render inside the conversation instead of a wall of text. This is what separates a ChatGPT app from a plain connector.

Users install the app from the directory, then call it with @YourApp or let ChatGPT suggest it when a request matches. Your app never sees the whole conversation, only the arguments ChatGPT passes to your tools.

How to build a ChatGPT app in 5 steps

1. Pick three to five jobs, not your whole product. Write down the sentences a user would actually type: "find a two-bedroom flat under $900 near a metro", "where is my parcel". Each sentence becomes one tool. Apps that expose twenty tools confuse the model and fail review for being vague about what they do.
2. Wrap your existing API as an MCP server. You rarely need new business logic. The server translates each tool call into calls your backend already supports, returns compact structured data, and enforces what the AI may touch. Read tools ship first; anything that changes data or spends money gets a confirmation step.
3. Add UI only where it beats text. A list of five flats with photos and prices is better as a carousel. A delivery status is better as one sentence. Every component you add is code to maintain and a screen to review, so start with one or two.
4. Add sign-in if users need their own data. Anything personal (orders, bookings, account balances) needs OAuth so ChatGPT acts as that user and only that user. This is usually the single largest line in the budget, which is why read-only public data is the cheapest first app.
5. Test in developer mode, then submit. Business, Enterprise and Edu admins can enable developer mode to run private internal apps. For the public directory, the submission needs an icon, name and description, screenshots, a privacy policy, a support contact, a verified website and your MCP server details. The app must suit general audiences including teens, and OpenAI reviews one version at a time before it goes live.

What you can and cannot sell in a ChatGPT app

This is where expectations need adjusting. As of September 2026, OpenAI's own monetization docs list external checkout as the generally available route: the app hands the user a link and the purchase completes on your site. Only physical goods are allowed. Digital goods and subscriptions are not permitted inside apps yet, and in-chat payment through Instant Checkout (built with Stripe on the Agentic Commerce Protocol) is a private beta limited to selected merchants.

So the realistic business case today is distribution, not a new till. A ChatGPT app puts your catalogue, booking flow or tracking page where people already ask the question, and sends qualified buyers to your checkout. That matters most for businesses whose customers start with a question rather than a search: travel, real estate, logistics, education, and stores with large catalogues. A cross-border shopping service like Meest Shopping, where a first-time customer's real question is "can I get this delivered home and what will it cost", is exactly the shape of product that answers well in chat.

How much it costs to build a ChatGPT app

Published 2026 guides put a focused MCP integration at $8,000 to $15,000 with a traditional team, and the enterprise end runs into six figures, driven mostly by authentication and the number of tools rather than the protocol code. An Apps SDK app adds the interface layer on top. Here is the same scope side by side:

Hosting the server runs $30 to $800 a month depending on traffic. The low prices come from fewer hours, not thinner work: AI writes the server, the components and the tests, and senior engineers spend their time on the tool list, the permission model and the review. That is how the same working app comes out 3 times cheaper and 3 times faster than a hand-written build.

How we would build yours at DForce

We already run MCP in production: our internal AI Brain serves the team's knowledge to every agent we use through an MCP server, so the tool design and permission patterns are ones we use daily, not ones we read about. For a ChatGPT app we start with the one-page list of sentences your customers would type, cut it to the three that lead to revenue, ship read-only tools and one card first, and submit to the directory once the app behaves on real questions, including hostile ones.

If your customers already ask ChatGPT about your category and your product is not in the answer, book a discovery call and we will tell you whether an app, a plain MCP connector, or neither is the right move for you.

Frequently asked questions

What is a ChatGPT app?

A ChatGPT app is a third-party product that runs inside a ChatGPT conversation, built with OpenAI's Apps SDK. It has an MCP server that gives ChatGPT tools to call your backend and an optional interface (a card, list, map or form) that renders in the chat. Users find it in the app directory and call it with @ and its name.

How do you build a ChatGPT app?

Pick three to five user jobs, wrap your existing API as an MCP server with one tool per job, add in-chat UI only where it beats text, add OAuth sign-in if users need their own data, then test in developer mode and submit to the directory. OpenAI reviews every submission before it goes live.

Can you make money with a ChatGPT app?

Yes, within limits. As of September 2026 the generally available route is external checkout on your own site, for physical goods only. Digital goods and subscriptions are not allowed inside apps yet, and in-chat Instant Checkout is a private beta. Most businesses use an app as a channel that sends buyers to their own checkout.

How much does it cost to build a ChatGPT app?

With an AI-first team, a focused app is $2,000 to $4,000 in 1 to 2 weeks, a production app with sign-in and write actions is $5,000 to $12,000 in 2 to 4 weeks, and a commerce or multi-tenant app is $12,000 to $25,000 in 4 to 8 weeks. That is 3 times cheaper and 3 times faster than traditional hand-written development.