ChatGPT Apps for Business: How to Put Your Software Inside ChatGPT (2026)

At DevDay on 29 September OpenAI expanded plugins into apps that live inside ChatGPT, with panels, file viewers and event triggers. Under the hood it is still an MCP server. Here is what changed, what it takes to build one, and why the same work also gets you into Claude.

Hugo Matthaey

Hugo Matthaey

Oct 16, 2026

A chat window frame with application cartridges slotting into it, a sidebar, a document viewer and a chart, all wired to a small company server behind it

ChatGPT apps are applications that run inside ChatGPT: a user calls them in a conversation to look something up, take an action, or open an interface, without leaving the chat. At DevDay on 29 September 2026, OpenAI expanded them: apps can now add sidebar experiences, interactive panels and custom file viewers, and start automations when something happens in a connected application. Underneath, they are built with OpenAI's Apps SDK, which OpenAI says is built on the Model Context Protocol (MCP). That matters for a business: the work of putting your software inside ChatGPT is mostly building a good MCP server, and the same server can put you inside Claude too. Below: what changed, how an app is built, how to get listed, and what we would do first.

What changed at DevDay 2026

From InfoQ's DevDay recap and CNBC's live coverage:

  • Plugins became apps. Developers can build "sidebar experiences, interactive conversation panels, and custom file viewers". CNBC described them as "essentially entire applications that feel native to ChatGPT".
  • Event triggers. Plugins can start automations "when an event occurs in a connected application", through what OpenAI calls the proposed MCP Events specification. Proposed means not final, so build on it knowing it may change.
  • Persistent agents. OpenAI introduced Dots, agents that keep working on ongoing tasks. Availability details were thin at launch.
  • Computer use in the Agents API, letting applications operate software through its graphical interface.

The direction is clear. ChatGPT wants to be where work starts, and your software either shows up there or gets copied and pasted into it.

What a ChatGPT app is, under the hood

Three layers:

  1. Tools. The actions your app exposes, such as "find an order", "create a quote", "update a ticket", each with a name, a description the model reads, and an input schema. This is an MCP server.
  2. Identity. Each user signs in with their own account in your application, through OAuth, so the app acts with their permissions and nobody else's.
  3. Interface (optional). Panels, viewers and widgets rendered inside ChatGPT with the Apps SDK. Useful when a result is better shown than described, like a quote to review or a document to check.

OpenAI made the Apps SDK open source when it launched apps in ChatGPT, and said apps built with it "can run anywhere that adopts this standard" (VentureBeat). The tools and identity layers are plain MCP. Claude connects to the same kind of server as a custom connector, so one server can serve both.

How to get listed

OpenAI opened app submissions on 17 December 2025, through the OpenAI Developer Platform. A submission includes:

  • the MCP connection details of your server;
  • testing instructions for the reviewers;
  • directory metadata for the listing;
  • country availability.

Apps have to pass OpenAI's quality and safety review, follow its usage policies, be suitable for general audiences including teens, and include a privacy policy and a support contact. At the time, selling digital goods, subscriptions or in-app services inside ChatGPT was not allowed, and purchases of physical goods had to link out. Check the current rules on OpenAI's plugins help article before planning revenue around it.

You do not have to be in the public directory. An app for your own staff, or your own customers who sign in, can be connected privately. For most B2B software, that is the first step anyway.

What we would build first

Not an interface. The tools.

  1. Pick the three actions your users copy into ChatGPT today. Look at support tickets and sales calls: "check an order status", "draft a quote", "summarize this account". Those are your first tools.
  2. Make each tool narrow, typed and honest about what it changes. Mark read-only tools as read-only, and treat everything else as a write that needs a stronger permission.
  3. Use your users' own sign-in. OAuth with the account they already have, and roles read from your database on every call, so access ends when their account does.
  4. Ship to your own users first, privately, and watch which tools the model actually calls and where it gets confused.
  5. Then add an interface where a result needs to be seen, and submit to the directory if public distribution is the goal.

Steps 1 to 4 are an MCP server. We show what that looks like in code, and the production fixes that matter, in how to build an MCP server.

What we do

We build and run MCP servers that put a company's software inside Claude and ChatGPT: typed tools over your real data, sign-in with your users' existing accounts, read and write permissions per tool, and roles checked against your database on every call. One server, both assistants.

Related: how to build an MCP server, Claude connectors explained, MCP security in 2026, and the Personal Agent Protocol.

Read next: Microsoft 365 Copilot pricing, Agentforce pricing and the Claude route.

Also: OpenAI Agent Builder shutdown and where to move.