phidea
← ChatGPT nano build guide
Reference · page 1 / 6

# 1. Overview — what "an App in ChatGPT" actually is, and when to pick it

Part 1 of 6. Index · Next → Architecture

Apps SDK (the thing this series is about)

An "App in ChatGPT" is an MCP server plus a set of custom UI widgets that render inline inside the ChatGPT conversation.

  • The model calls your MCP tools the same way it calls first-party tools.
  • Your server returns a JSON result for the model and an HTML/JS bundle that ChatGPT renders in a sandboxed iframe.
  • Think "interactive response card backed by your backend," not "chatbot persona."
  • Works across ChatGPT web and mobile from a single server.

Pick this when you need your UI, your data, and your backend logic to live inside the chat surface.

vs. Custom GPTs / GPT Store

No-code personas built in the ChatGPT UI. You give them a system prompt, optional knowledge files, and optional Actions (OpenAPI-defined HTTP calls). They cannot render custom UI, cannot be embedded outside ChatGPT, and are best for internal hacks or personal assistants.

An Apps SDK app is a superset: arbitrary React UI, cross-turn state, discoverable through the new in-product app directory rather than the GPT Store.

vs. AgentKit / Agent Builder / ChatKit

Announced at the same DevDay (Oct 2025). AgentKit is for building agentic workflows you embed in your own product (via ChatKit) or trigger from the Responses API.

  • Apps SDK = "my app lives in ChatGPT."
  • AgentKit = "an OpenAI agent lives in my app."

They can coexist — your MCP server can back both.

vs. Assistants API

The older (2023–2024) stateful thread+tools API. OpenAI has guided new work toward the Responses API and, for chat-native UX, toward Apps SDK. Use Apps SDK instead if your goal is to ship something users discover in ChatGPT.

When to pick Apps SDK

All of the following should be true:

  • You want distribution inside ChatGPT.
  • You need rich interactive UI (not just text).
  • You already have a backend/API you can expose as MCP tools.
  • You are OK running a production MCP server (HTTPS, uptime, logging).

If any of those are false, reconsider:

  • No backend → Custom GPT with Actions is lighter.
  • No chat distribution needed → AgentKit embedded in your own product.
  • Just a system prompt + knowledge base → Custom GPT.