OpenAI publishes MCP Extensions, an Apache-2.0 SDK that lets an MCP plugin add a ChatGPT sidebar, file viewer and composer mentions

Building a plugin for ChatGPT has meant losing every UI detail a built-in feature has. OpenAI published mcp-extensions on 29 September, an SDK that gives an MCP plugin four ChatGPT-specific features from the same server that already talks to Claude, Cursor and Codex. The repository is Apache-2.0 and, at the time of writing, holds 266 stars.

The four features OpenAI names in the README, and what each one adds:

• Sidebar entrypoints. A plugin can register an entry in the ChatGPT left rail. A user selects it and opens a plugin surface directly, without going through a message.
• File extension handlers. When a user opens a file whose type the plugin supports, ChatGPT renders the plugin's own viewer rather than a generic preview. The showcase example opens a CAD file in a Bits & Bolts viewer.
• Composer at-mentions. A plugin's resources become searchable from the same at-mention menu that already picks people and files, so a user can add a reference from the plugin into the message they are writing.
• Extended forms. Forms rendered inside ChatGPT can carry richer inputs than plain text. The example the SDK ships with lets a user pick a CAD part from a grid of thumbnails.

What OpenAI shipped, and what still lives elsewhere

The repository holds the SDK, the specification, and the Bits & Bolts sample plugin. The packages are on the public registries: @openai/mcp-extensions for TypeScript, openai-mcp-extensions for Python. Both target an MCP server; the TypeScript package also targets Apps. The plugin distribution and directory listing stay in ChatGPT: OpenAI links to a plugin-directory install of Bits & Bolts Remote as the try-it-first step, and points at its existing plugin-building documentation for the plugin itself. The SDK is the piece that lets a plugin adopt these four ChatGPT-specific UI hooks; the plugin itself is a separate thing you already had to build.

The read for a team already running an MCP server

A team that has an MCP server for internal knowledge, a design system, a parts catalogue, or a data query tool has been able to expose that server to any MCP-aware client for months. What changed on 29 September is that the same server can now light up four features specific to one client, ChatGPT, without a second implementation. A parts catalogue can appear in the sidebar, render its own CAD viewer, be at-mentioned from the composer, and let a user pick a part from thumbnails rather than typing a SKU. Nothing about the server for Claude or Cursor has to change.

The trade is that these hooks are ChatGPT-specific by construction and not part of the MCP standard, so a plugin that uses them is doing extra work for one destination. That is the same shape as building a Claude Skill or a Cursor integration: a portable base with a client-specific top layer. The SDK is Apache-2.0 and the specification is in the same repository, so a competing client could implement the same hooks. Whether any of them will is not something the release itself answers.

Source

• Primary: openai/mcp-extensions on GitHub, Apache-2.0
• Specification: mcp-extensions spec
• TypeScript SDK: @openai/mcp-extensions on npm
• Python SDK: openai-mcp-extensions on PyPI