Plugins vs apps vs MCP servers vs skills: what's the difference?
How ChatGPT plugins, apps, MCP servers and skills relate, with a comparison table and a quick guide to which one you need as a user or a builder.
An MCP server is the service that exposes tools and data to an AI client. An app is an MCP server that has been connected to ChatGPT, optionally with its own interface. A skill is a folder of instructions (a SKILL.md file) that teaches the model how to do a specific job. A plugin is the installable package that bundles any of these, and it is what you find in the ChatGPT and Codex Plugin Directory. Most confusion comes from the fact that these are layers, not alternatives: a plugin often contains an app, which is built on an MCP server, plus a few skills.
The rest of this post defines each one, shows how they fit together, and helps you decide which one you need.
The four concepts at a glance#
| MCP server | App | Skill | Plugin | |
|---|---|---|---|---|
| What it is | A server speaking the Model Context Protocol | An integration in ChatGPT, built on an MCP server | Instructions and resources for one kind of task | An installable bundle |
| What it contains | Tools, auth, optional UI resources | MCP server + optional UI + listing metadata | SKILL.md, plus optional scripts, references, assets | Skills, MCP servers/apps, app templates, hooks |
| Does it run code? | Yes, on a server or your machine | Yes, through its MCP server | Only if it ships helper scripts | Depends on what it bundles |
| Where it works | Any MCP client: ChatGPT developer mode, Codex, the OpenAI API, other agents | ChatGPT, and Codex through plugins | Codex, ChatGPT, the OpenAI API, other agents that support skills | ChatGPT (Chat and Work) and Codex |
| Who usually builds it | The service provider or a developer | The service provider | Anyone who can write Markdown | The provider, OpenAI, or a team |
| How you get it | Add a URL or command to your client | Install from the Plugin Directory | Install via Codex, a plugin, or copy a folder | Install from the Plugin Directory or a marketplace |
| Old name | n/a | Connector (until December 2025) | n/a | n/a (the 2023 "plugins" were a different system) |
MCP servers: the plumbing#
The Model Context Protocol is an open standard for connecting AI clients to tools and data. An MCP server lists the tools it offers (with names, descriptions and input schemas), runs them when the client calls them, and handles authentication, usually OAuth for hosted servers.
MCP servers come in two transport flavours:
- Remote (Streamable HTTP): hosted at a URL such as
https://mcp.linear.app/mcp. This is what ChatGPT connects to. - Local (stdio): a command that runs on your machine, such as the Playwright browser-automation server. Codex and the ChatGPT desktop app can run these; ChatGPT on the web cannot.
An MCP server is not tied to ChatGPT. The same server works in Codex, in the Responses API mcp tool, and in other MCP clients. Our MCP directory lists more than 140 of them.
To connect one yourself, see Add an MCP server to ChatGPT or Use MCP in the OpenAI API.
Apps: an MCP server inside ChatGPT#
An app is how an MCP server shows up for ChatGPT users. It adds a listing (name, description, icon), goes through OpenAI's review if it is public, and can render an interface in the conversation: a map of hotels, a playlist, a design preview.
Custom UI is optional. ChatGPT implements the open MCP Apps UI standard, where a tool points to a UI resource with _meta.ui.resourceUri, and adds ChatGPT-only extras through window.openai, such as checkout, file uploads and modals (OpenAI docs).
Before December 17, 2025, data-only integrations such as Gmail or Google Drive were called connectors. OpenAI renamed connectors to apps so one word covers both. And since July 9, 2026, every app in the directory is packaged inside a plugin (details).
If you want to build one, start with Build your first ChatGPT app and test it with MCP Inspector.
Skills: instructions the model loads when needed#
A skill is a folder containing a SKILL.md file with YAML front matter (name and description) and Markdown instructions. It can also carry scripts/, references/ and assets/ folders (OpenAI skills docs). Skills follow the open Agent Skills format, so the same folder can work across agents that support it.
Skills use progressive disclosure: at first the model sees only each skill's name and description. It loads the full instructions only when a task matches. That keeps dozens of installed skills from crowding the context.
A skill does not connect to anything by itself. It tells the model how to do a job (your team's code review checklist, how to format a board report, which Stripe APIs to prefer) using tools it already has. In Codex you invoke one with $skill-name; in ChatGPT, with @.
Want to write one? See Write a skill for ChatGPT and Codex.
Plugins: the package you install#
OpenAI describes plugins as "the packages people discover, install, share, and publish in ChatGPT and Codex" (plugin architecture). A plugin can contain:
- Skills for repeatable workflows
- MCP servers / apps for tools and data, with optional UI
- Hooks that run commands at set points in the Codex runtime
- App templates, according to OpenAI's help article
The docs name four common shapes: skills only, MCP server only, skills plus MCP server, and MCP server with UI. The portable manifest is a plugin.json file at the plugin root. ChatGPT and Codex share one directory, and teams can also publish plugins through their own marketplaces.
One product, three layers: Figma#
Figma is a good example because it exists at every layer in our catalog.
The MCP server is Figma's hosted server at mcp.figma.com. You can add it to Codex or any MCP client to pull design context into code.
The app is Figma's ChatGPT integration: create FigJam diagrams and design files from a conversation.
The plugin bundles the Figma app connection with Figma's skills for design-to-code, Code Connect and design system work, so the model knows how to use the tools well, not only that they exist.
Not every plugin has an app. Superpowers, a community plugin in the curated Codex marketplace, is skills only: planning, test-driven development and debugging methods, no external service.
Which one do I need?#
"I want ChatGPT to work with a service I use." Look for a plugin in the Plugin Directory. Install it and connect your account. See How to install plugins in ChatGPT.
"The service has an MCP server but no plugin." Add the MCP server yourself. In ChatGPT that needs developer mode on an eligible plan; in Codex it is one codex mcp add command. See Add an MCP server to ChatGPT. Read the security checklist first: developer mode includes write tools.
"I keep explaining the same process to ChatGPT or Codex." Write a skill. No server or code needed.
"I'm building an integration for my product." Build an MCP server, add UI only if it helps, add skills for your main workflows, and package it as a plugin. Then publish it.
"I'm a developer working in Codex." Use plugins from the shared directory or a marketplace, and add standalone MCP servers in ~/.codex/config.toml. See Use plugins in Codex and our best plugins for developers.
"I'm calling the OpenAI API, not using ChatGPT." Plugins don't apply. Pass remote MCP servers to the Responses API mcp tool, or mount skills in the hosted shell. See Use MCP in the OpenAI API.
Common mix-ups#
- "ChatGPT plugins were discontinued." The 2023 plugins were, in April 2024. The current plugins are a new system that started in Codex in March 2026 and took over the ChatGPT directory in July 2026. Full story: What happened to ChatGPT plugins?
- "An app and an MCP server are the same thing." Close, but an app adds a ChatGPT listing, review and optional UI. Many MCP servers in our /mcp directory have no ChatGPT app.
- "Skills are just prompts." They are closer to playbooks. They can include scripts and reference files, and the model loads them only when relevant.
Note (October 4, 2026): DevDay 2026 added two pieces that sit on top of these layers. Plugin Extensions let a plugin's app appear as a sidebar app, conversation panel, file viewer or editor, or rich form. MCP Events let an MCP server push updates that ChatGPT can act on. See DevDay 2026: what it means for plugin builders.
