Building a personal app inside ChatGPT? Have your Dot host the backend on a Site and ship a plugin

#What changed

Simon Smith (@_simonsmith) says he has now built two personal apps that live inside ChatGPT: a customizable news feed with thumbs-up/down feedback, and a Spanish vocabulary trainer that imports Kindle highlights into flash cards and quizzes. His recipe was to tell his Dot to build the backend as an MCP server hosted on a ChatGPT Site and the frontend as a plugin with extensions, so the app docks in the ChatGPT sidebar.

OpenAI's help center documents the hosting half of this: you can ask ChatGPT or Codex to add an MCP server to a new or existing Site, and once the Site owner publishes it, ChatGPT creates a plugin you install and connect. Plugin extensions are what put a plugin in the sidebar, the composer, and file viewers.

His main observation about Dots: they act as a build babysitter. The same vocab app built directly in a Work thread ended up with a broken connection between plugin and MCP server, while his Dot kept testing and fixing until it worked.

#Who is affected

ChatGPT users with Dots who want small personal tools (feeds, trackers, study apps) that run inside ChatGPT instead of a separate web app. Site-hosted plugins are available on all plans per OpenAI; Business and Enterprise members need an admin to turn on the plugin and Site publishing permissions first.

#What to do now

• Ask your Dot for the architecture by name: "Build this with a Sites-hosted MCP server and a plugin with extensions." He says the Dot figures out the rest.
• In the same request, ask for proper sidebar icons and a plugin icon that matches them. Without that, his Dot left the default "@" icon.
• Ask it to follow ChatGPT's own styling unless you want a custom look, so the app feels native.
• After any change to the tools, the Site owner has to publish the Site again before the plugin picks up new actions, per OpenAI's help article. Then check the plugin's tool list and try one action.
• Do a final hands-on test yourself. His Dot sometimes could not fully test the plugin, once because of Sites sign-in trouble.

#What is not confirmed

This is one user's field report, not independently verified. The plugin version label staying at v1.0.0 even after the code changes is his observation and may be a display bug or a gap in the Dot's build skills; OpenAI has not confirmed either. His comparison with Muse (faster, better at intent) and Claude artifacts (similar first result, less native integration) is his own experience. There is still no built-in way to see what a Dot is doing mid-task; he asked OpenAI for a receipt reaction, a short status line, and a view of the threads a Dot is overseeing.