OpenAI wants your MCP plugins to stop looking like guests

What it does

This repo extends MCP with ChatGPT-specific capabilities so a plugin can behave less like a bolt-on tool and more like part of the app. Concretely: your plugin can register a sidebar entrypoint, take over file viewing for file types it supports, surface its resources as @-mentions in the composer, and render richer elicitation forms (e.g. thumbnail pickers). There’s a TypeScript SDK (@openai/mcp-extensions) for MCP servers and Apps, plus a Python one (openai-mcp-extensions) for servers.

The interesting bit

MCP standardizes how models call tools, but says nothing about UI — so plugins end up as chat-only oracles. These extensions fill that gap with actual surface area inside ChatGPT, which is a meaningful shift in what “plugin” means here. The spec lives in docs/spec.md, and the bundled Bits & Bolts CAD demo plugin shows the whole thing working end to end.

Key highlights

• Sidebar entrypoints so users reach your app without prompting
• Custom file viewers triggered by file extension
• Composer @-mentions that search and reference your plugin’s resources
• Extended elicitation forms with image/thumbnail choices
• Reference plugin (Bits & Bolts) plus a patterns doc for best practices

Caveats

• These are ChatGPT-specific extensions, not part of the MCP standard — portability to other MCP clients is unclear from the README
• The README leans on asking @Plugin Creator in Codex to get started; the docs beyond the spec links aren’t deep

Verdict

Worth a look if you’re building MCP plugins and want them to feel native in ChatGPT rather than conversational-only. If you target multiple MCP clients, the vendor-specific layer deserves a careful read first.