OpenAI's MCP Extensions put plugins in the ChatGPT sidebar

Open Bits & Bolts Remote from the ChatGPT sidebar and a Parts Library appears, where you select a CAD part from thumbnail choices. The demo plugin shows off OpenAI MCP Extensions, a set of ChatGPT-specific additions to MCP that OpenAI has published on GitHub under the Apache License 2.0.

At a glance

• OpenAI has published four extensions to MCP that put a plugin in ChatGPT's sidebar, show custom file viewers, support composer mentions and render richer forms with thumbnail choices.
• Developers add the extensions to an existing MCP server through two SDKs: @openai/mcp-extensions for TypeScript servers and Apps, and openai-mcp-extensions for Python servers.
• The extensions are ChatGPT-specific, and the Python package covers MCP servers only, so the project lists Apps support for the TypeScript SDK alone.

If you have not been following, MCP is the open protocol Anthropic introduced so AI assistants can reach external tools and data through a common server interface. You write a server once, and any client that speaks the protocol can call it. ChatGPT plugins can be built on MCP servers, and OpenAI's plugin documentation explains how to create one.

The repository names four extensions. Sidebar entrypoints let users open your app straight from the ChatGPT sidebar. File extension handlers render a custom viewer when a user opens a file type your plugin supports. OpenAI's stated goal is plugins that feel like native, first-class features of ChatGPT rather than tools called from inside a chat.

The other two work inside the conversation. Composer mentions let users search your plugin's resources from the composer and add references to a message. Extended forms give richer inputs than plain text fields. In the Bits & Bolts demo, that means selecting a CAD part using thumbnail choices.

The repository numbers its getting-started steps from zero. Step zero is installing Bits & Bolts Remote from the plugin directory and selecting it in the sidebar to open the Parts Library. Step one is creating a plugin by following OpenAI's plugin documentation. After that, you add the SDK for your language to your MCP server and read the spec for the details of each supported extension.

Here is how the pieces fit together. Plain MCP gives a server and a client a shared vocabulary for tools and resources. The extensions add words to that vocabulary that only ChatGPT understands: where to place an entry in the sidebar, which file types to claim, what to show when someone searches from the composer. Think of a mouse with extra buttons that only work once you install the maker's software.

The repository leaves several things out. It does not list which file types the handlers support, and it does not say how other MCP clients treat a server that uses these extensions. In our view, it is odd that the Python SDK stops at servers when the TypeScript one also covers Apps, since plenty of MCP servers are written in Python.

Whether Python gets Apps support

The repository gives no timeline for bringing Apps support to the Python SDK, and it does not say whether more extensions are coming beyond the four listed. It also does not say whether other MCP clients will adopt any of them. For now, the spec in the repository is the document that defines what the supported extensions are.