From Apps SDK to ChatGPT Plugins
On this page
Course rewritten on November 3, 2026 based on OpenAI’s Plugins documentation. The previous version relied on a non-existent
@openai/apps-sdkpackage (featuringdefineAction,defineWidget, and achatgpt-app.jsonmanifest). You now build ChatGPT apps using the MCP protocol, its official SDKs, and the MCP Apps interface standard. We have rewritten all fourteen lessons and removed the unsupported model compatibility lists, commission structures, and publishing steps from the old version. Sources: Quickstart · Plugin architecture.
What is a plugin?
A plugin is the package you discover, install, share, and publish within ChatGPT and Codex. It extends the assistant’s capabilities by providing new knowledge, tools, or both. OpenAI’s documentation identifies four building blocks:
- Skills, which give the model instructions and resources to handle recurring workflows;
- An MCP server, which exposes tools and connects the assistant to external systems—your API, database, or service;
- Both, when the model needs both a playbook and capabilities served by your infrastructure;
- Lifecycle hooks, commands executed at specific moments in the Codex runtime environment, including within ChatGPT Work.
ChatGPT and Codex share a single plugin directory: a public plugin is published once and becomes visible across compatible surfaces in both products. Some capabilities remain surface-specific—for example, a hook script must exist in the runtime environment, so installing the plugin on the web does not automatically deploy it there.
What about the “Apps SDK”?
This term refers to OpenAI’s toolkit for building applications inside ChatGPT: the MCP protocol and its official SDKs, the open MCP Apps standard that ChatGPT implements to render interfaces, ChatGPT-specific extensions, a design system (@openai/apps-sdk-ui), and a repository of examples. The documentation now groups these under “ChatGPT plugins and Apps SDK”: an app centers on an MCP server, includes an optional interface, and distributes as part of a plugin.
The choice of MCP is deliberate. The Model Context Protocol is an open specification for connecting AI clients to tools and data. The server you write for ChatGPT speaks a protocol other hosts understand, and an interface compliant with MCP Apps works in any host supporting that standard.
Four possible forms
Start with the smallest form that covers your use cases; you can add a server or interface later without changing the plugin’s core purpose.
| Form | Choose this when |
|---|---|
| Skills only | Existing instructions and tools are sufficient to handle the workflow |
| MCP server only | The plugin needs MCP tools but no additional workflow instructions |
| Skills and MCP server | Skills must guide the model through workflows that use your tools |
| MCP server with interface | Visual interaction significantly improves part of the workflow |
A plugin that drafts meeting minutes according to your company rules might only need skills. As soon as it must read live data, authenticate a user, or act within a system you operate, it requires an MCP server. Finally, an interface is never mandatory: a tool must remain useful without one.
Where the plugin runs
In ChatGPT, you invoke an installed plugin directly by typing @ in the input field and selecting it. The model may also call it automatically when a request matches its tools or skills. OpenAI’s quickstart guide recommends testing your first plugin in ChatGPT Work, the Work tab on the ChatGPT web homepage. The same plugins appear in Codex, limited by the specific capabilities of each surface.
What you will build
This course follows the path outlined in the documentation. You will start by defining your tools based on user needs, connect a first MCP server to ChatGPT, and then build your own: well-documented tools, an MCP Apps interface where it adds value, OAuth authentication for private data, and, if your business model fits, a purchase flow using the Checkout API. You will finish with packaging, publishing rules, and submission to the directory.
Key takeaways
- A plugin combines skills, an MCP server, or both; the interface is optional
- ChatGPT and Codex share a single plugin directory
- “Apps SDK” refers to OpenAI’s tools for building apps: MCP, the MCP Apps standard, ChatGPT extensions, and the design system
- There is no
@openai/apps-sdkpackage: start with the official MCP SDKs - Choose the smallest plugin form that covers your use cases