ChatGPT Apps vs Plugins: What Changed in 2026
A plugin is the installable package; an app is the MCP-backed integration inside it that connects to a real service and does the work. They are not competing names for one thing — they name different layers, and both are current. In July 2026 OpenAI migrated the App directory to the Plugin directory, which changed the packaging and the vocabulary, not what the integrations do.
The confusion is understandable, because the word plugin has been reused. It meant something different in 2023, that generation was retired, and the name has now come back attached to a different structure. This page separates the layers so you can tell which one a given article, tutorial, or search result is talking about.
The layers, in one table
Six words get used interchangeably in ChatGPT coverage. They are not interchangeable. Here is what each one actually names.
| Layer | What it names | What it contains | Example | Ranked here? |
|---|---|---|---|---|
| Plugin | The installable package people discover, install, and share | Skills, an MCP server, or both | The listing you install from the plugin directory | Yes — when it contains an MCP server |
| App | The MCP-backed integration that does the work in a conversation | Tools, auth, structured results, optional UI | Canva, Notion, Spotify, Wolfram | Yes — this is what we score |
| Skill | Packaged instructions for a repeatable workflow | A SKILL.md file, plus optional scripts and templates | A meeting-recap workflow bundled in a plugin | No — different product shape |
| MCP server | The backend that exposes tools and connects to a service | Tool schemas, authentication, optional UI resources | The server behind the Notion plugin | Indirectly — it is what we test |
| Connector | The account link granting ChatGPT access to your data | OAuth scopes on your account | Your Google Drive authorization | Only when surfaced as an app |
| GPT | A custom assistant built in the GPT builder | Instructions, files, optional actions | A Python tutor GPT | No — see apps vs GPTs |
How a plugin is put together
OpenAI's developer documentation defines plugins as the packages people discover, install, share, and publish in ChatGPT and Codex. Its user-facing documentation goes further and lists six kinds of component a plugin can bundle — most coverage mentions two, which is why the format gets described as narrower than it is.
| Component | What it contributes |
|---|---|
| Skills | Reusable instructions for a kind of work, loaded when relevant so the model follows the right steps and uses the right references. |
| Connectors | The connection you authorise to a tool like GitHub, Slack, or Google Drive, so ChatGPT can read from it and act in it. Can include custom UI. |
| MCP servers | The services behind connectors. They define tools, enforce auth, return structured data, and perform actions against external systems. |
| Browser extensions | Browser capabilities a plugin's workflow depends on. |
| Hooks | Commands that run at configured lifecycle points. OpenAI's documentation says to review and trust a plugin's hooks before enabling them. |
| Scheduled task templates | Reusable starting points for recurring tasks, where scheduled tasks are available. |
The connector-and-MCP-server pair is the one most worth getting straight, because they are not alternatives: an MCP server is the service behind a connector. "Connector" is the word the product uses for the connection you authorise; "MCP server" is the word the developer docs use for the thing serving it. Same layer, two audiences.
Plugin
├── Skills → workflow instructions the model loads on demand
├── Connectors (optional) → the connection you authorise (Gmail, Slack, Drive)
│ └── backed by an MCP server → tools, auth, structured results, optional UI
├── Browser extensions → browser capability a workflow needs
├── Hooks → commands at lifecycle points — review before enabling
└── Scheduled task templates → starting points for recurring tasksHooks are worth a second look before you install anything: they run commands at configured lifecycle points, and OpenAI's own documentation says to review and trust a plugin's hooks before enabling them. That is a materially different risk profile from an app that reads your calendar.
Where plugins actually work
Surface availability is documented and frequently misreported. Straight from OpenAI's user documentation:
| Surface | Plugins? | Notes |
|---|---|---|
| ChatGPT on the web | Yes | Chat and Work. |
| ChatGPT desktop app | Yes | Chat, Work, and Codex. |
| ChatGPT on mobile | Yes | Plugins available to your account, in Chat or Work. |
| Codex CLI | Yes | A separate plugin browser, opened with /plugins. |
| IDE extension | No | Documented as unsupported — use the desktop app or Codex CLI instead. |
The Plugins Directory itself is organised into tabs — plugins built by OpenAI, plugins provided by your workspace, and a Personal tab covering ones you created or that were shared with you — plus a separate row for what you have already installed. That structure is part of why published counts of "how many plugins exist" disagree so widely: they are not all counting the same scope.
The three shapes that matter for ranking
Six component types produce many possible combinations, but only one distinction decides whether we can score something: is there a service behind it?
| Shape | MCP server | Custom UI | What it means for you | In our rankings |
|---|---|---|---|---|
| Skills-only plugin | None | None | Packaged instructions and templates. Nothing connects to an outside service, so nothing can rate-limit, break, or ask for your data. | Not scored |
| MCP-only plugin | Yes | None | A real integration that returns structured data and model-readable text without a custom interface. Fully functional; it just renders as conversation. | Scored |
| MCP + UI plugin | Yes | Yes | The shape most people picture: an integration that draws something you can inspect, compare, edit, or confirm inside the chat. | Scored |
What actually changed in July 2026
The directory was renamed and re-scoped. The App directory became the Plugin directory. ChatGPT and Codex now share one universal directory, so publishing a public plugin produces a single listing that resolves from supported surfaces in either product. Individual capabilities can still be surface-specific — a plugin can include hooks that only run in Codex.
Skills became a first-class distributable. Before, the unit you published was an integration. Now you can publish a package that contains only workflow instructions, only a server, or both. That is the substantive change, and it is why the packaging needed a new name.
Nothing broke for users. Existing app connections carried over. You add new plugins from the directory and authenticate the underlying service as before.
The old vocabulary is still visible in the product. The Apps SDK kept its name. The permission gating plugin submission on the OpenAI Platform is still labeled "Apps Management." OpenAI's own docs use "app" and "plugin" side by side. If you were waiting for one term to win, it has not happened, and the layered reading above is more useful than picking a side.
Vocabulary translator
What a term meant, what it means now, and whether you can trust what you are reading.
| If you see | It means today | Note |
|---|---|---|
| “ChatGPT plugin” (2023 articles) | Retired generation — steps no longer apply | Manifest-based, own store, per-chat toggle. Gone. |
| “ChatGPT plugin” (mid-2026 onward) | The installable package | Can contain skills, an MCP server, or both. |
| “App directory” | Plugin directory | Migrated July 2026. Existing connections kept working. |
| “ChatGPT app” | Still current | The MCP-backed integration. The Apps SDK kept its name. |
| “App store” | Never existed under that name | The directory is the closest thing. |
| “Apps Management” permission | Still called that | The permission gating plugin submission on the Platform. |
What this site ranks, stated precisely
We rank MCP-backed plugins that a person can invoke inside a ChatGPT conversation — the layer this site has always called an app. The migration did not change our scope; it gave us a more precise way to state it. Canva, Notion, Wolfram, and Spotify are in scope whether you call them apps or plugins.
We do not score skills-only plugins. Our seven criteria — usefulness, reliability, ease of use, setup, in-chat experience, privacy clarity, and value — exist to compare integrations that connect to a real service and can therefore break, rate-limit, or expose data. For a folder of instructions with no service behind it, most of those dimensions have no meaning. Publishing a score anyway would mean publishing a number with nothing underneath it, and that is the one thing this site does not do. If skills-only listings become a large enough share of the directory to deserve comparison, they will get their own criteria and their own surface — not a borrowed score.
Custom GPTs remain out of scope for separate reasons, covered on apps vs GPTs. Read the ranking methodology for how the seven criteria are weighted.
Frequently asked questions
Frequently asked questions
- Are ChatGPT apps and plugins the same thing?
- Not quite — they name different layers. A plugin is the installable package people discover and install; an app is the MCP-backed integration inside it that connects to a real service and does the work. Most listings you would call an app are distributed as a plugin. Both terms are current: OpenAI's docs use both, and the Apps SDK kept its name.
- Did ChatGPT plugins come back?
- The word did, but not the product. The 2023 plugin generation — manifest-based, its own store, a per-chat enable toggle — was retired in favor of apps and connectors. In July 2026 OpenAI migrated the App directory to the Plugin directory, and “plugin” now means the installable package that bundles skills and an MCP server. A tutorial written for 2023 plugins will not work.
- What is inside a plugin?
- Six possible component types, not two: skills (reusable workflow instructions), connectors (the connection you authorise to a tool like Slack or Drive), MCP servers (the services behind connectors), browser extensions, hooks (commands that run at lifecycle points — OpenAI says review and trust them before enabling), and scheduled task templates. Most coverage mentions only skills and MCP servers, which makes plugins sound narrower than they are.
- Is a connector the same as an MCP server?
- They are the same layer described to two audiences. A connector is the connection you authorise; an MCP server is the service behind it, defining tools, enforcing auth, and returning structured results. The product says connector, the developer docs say MCP server.
- Do plugins work everywhere ChatGPT does?
- Almost. They work in Chat and Work on web, desktop, and mobile, and in Codex in the desktop app, and Codex CLI has its own plugin browser. The IDE extension does not support plugins at all — documented, and frequently misreported.
- Do I need to reinstall my apps after the directory migration?
- No. Existing app connections were not broken by the migration. You add new plugins from the plugin directory, authenticating the underlying service the same way you did before.
- Is the plugin directory the same in ChatGPT and Codex?
- Yes — they share one universal plugin directory, so a public plugin resolves to the same listing from supported surfaces in either product. Individual capabilities can still be surface-specific: a plugin can include hooks that only run in Codex.
- Why don't you rank skills-only plugins?
- Our seven criteria are built to compare integrations that connect to a real service and can therefore break, rate-limit, or expose data. For a folder of instructions with no service behind it, most of those dimensions are undefined. We would rather leave it out of the rankings than publish a score with nothing underneath it.
- Should I search for “best ChatGPT apps” or “best ChatGPT plugins”?
- Either — you are looking for the same thing, and we rank it under both names. Use plugin when you mean the thing you install, and app when you mean the integration that does the work.
Primary sources
Every definitional claim on this page traces to OpenAI's own documentation. Verified 2026-08-24.
- Plugin architecture — OpenAI Developers
Defines plugins as the packages people discover, install, share, and publish, and sets out what a plugin can contain.
- Skills — OpenAI Developers
Defines a skill and the boundary between a skill and an MCP server.
- Submit plugins — OpenAI Developers
Confirms the three submittable shapes and the Apps Management permission.
- Plugin guidelines — OpenAI Developers
Describes the minimum bar for staying in the universal directory and eligibility for enhanced distribution.
- Plugins (user documentation) — OpenAI
The six component types, surface availability, the directory's tab structure, and the connector-to-MCP-server relationship.
- Plugins in ChatGPT and Codex — OpenAI Help Center
The App directory to Plugin directory migration, dated July 9, 2026, and its effect on existing connections.