This comparison is part of our MCP comparisons, within The MCP Guide. Facts are checked against each vendor's own documentation and pricing pages on the date shown.
Build a plain MCP server if you want one integration that works in Claude, Cursor, ChatGPT and other AI apps. Package it as a ChatGPT plugin, the product that grew out of the Apps SDK, when you want a public listing in ChatGPT and Codex. You rarely choose one over the other, because an OpenAI plugin is an MCP server plus packaging, optional UI and a review process.
Checked against: OpenAI's Plugins documentation, the MCP Apps overview and the 2026-07-28 MCP changelog, read on 28 October 2026.
A naming note first. The old developers.openai.com/apps-sdk address now redirects to the Plugins docs. OpenAI's help center still says "Pro users can build apps using the Apps SDK," so both names are in use. This page uses "Apps SDK" for the developer toolkit and "plugin" for what ships.
Key Takeaways
- MCP is an open protocol any AI app can implement. The Apps SDK, now documented as Plugins, packages MCP servers for ChatGPT and Codex.
- OpenAI says a plugin can contain skills, an MCP server, or both, and that "ChatGPT and Codex share one universal plugin directory."
- For interactive UI, OpenAI says ChatGPT implements the open MCP Apps standard, so the same UI can run in other compatible hosts.
- Public plugins go through verification, automated scans and human review; a plain MCP server needs none of that.
- Custom MCP servers in ChatGPT are available to Business and Enterprise/Edu workspaces on the web, with read and fetch only on Pro.
At a glance
| Plain MCP server | ChatGPT plugin (Apps SDK) | |
|---|---|---|
| What it is | An open protocol for AI apps to call tools and read data | A package of skills, an MCP server, or both |
| Who maintains it | The MCP project under the Agentic AI Foundation | OpenAI |
| Where it runs | Any MCP client, local or remote | ChatGPT and Codex |
| Interactive UI | Optional, through the MCP Apps extension | Optional, MCP Apps plus window.openai extras |
| Public listing | MCP Registry and third-party directories | ChatGPT and Codex plugin directory |
| Review before listing | None required by the protocol | Verification, automated checks and human review |
| Commerce | Not covered by the protocol | Physical goods only, per OpenAI guidelines |
| Cost to build | Free and open source | No submission fee listed in the docs we read |
Where plain MCP is stronger
Reach beyond one vendor. MCP is not tied to a single AI company. The extension support matrix marked MCP Apps support in Claude (web), Claude Desktop, ChatGPT, Cursor, VS Code GitHub Copilot and others on 28 October 2026. One server and one UI can reach all of them. See AI apps that support MCP for the full list.
No gatekeeper. You can publish a remote server and share its URL the same day. No verification, review queue or policy attestations stand between you and your users. That suits internal tools and early pilots.
Fewer content rules. OpenAI's plugin guidelines say plugins "may conduct commerce only for physical goods" and must not sell subscriptions, tokens or credits. They also bar ads and require suitability for users aged 13 to 17. A plain MCP server you run for your own team carries none of these directory rules. Your own policies and the spec still apply.
Local servers. The core spec supports stdio servers that run on the user's computer. OpenAI handles local servers as a separate "desktop-only" submission path. Our stdio vs Streamable HTTP comparison explains the trade-off.
Where the ChatGPT Apps SDK is stronger
Distribution inside ChatGPT and Codex. OpenAI's quickstart says public plugins "are published once and become discoverable from supported surfaces in both products." If your buyers live in ChatGPT, that listing is the main reason to build one.
Skills and server in one package. OpenAI's plugin architecture page lists four shapes: skills only, MCP server only, skills plus server, and server with UI. Skills add written workflows around your tools. That bundle is one installable item for the user.
Extra UI features in ChatGPT. OpenAI tells developers to start with the shared MCP Apps standard. Its UI guide then lists ChatGPT-only extensions on window.openai, such as requestCheckout, uploadFile, requestModal and widgetState. Older Apps SDK fields like _meta["openai/outputTemplate"] remain as "compatibility aliases," but new UI should use _meta.ui.resourceUri.
Ongoing updates without resubmitting. OpenAI's submission guide says it scans your hosted MCP server daily after publication. Eligible tool changes "become available after automated checks pass," with no new package upload.
How the MCP Apps extension fits
MCP Apps is the bridge between the two. The MCP Apps overview says apps let servers "return interactive HTML interfaces" that render in the chat. A tool points to a ui:// resource, and the host renders it in a sandboxed iframe. Our MCP Apps extension guide covers the details.
OpenAI's docs say ChatGPT "implements the open MCP Apps standard." So the UI you build for a ChatGPT plugin can run in other MCP Apps hosts, as long as you avoid ChatGPT-only extensions or feature-detect them.
Pricing
MCP: the protocol, specification and official SDKs are free and open source.
ChatGPT plugins: the Plugins docs we read describe verification, review and publication steps but list no submission fee. Your users still need a ChatGPT plan that supports plugins.
Custom MCP servers in ChatGPT: OpenAI's help article on custom MCP servers says "Custom MCP server connections and full MCP support are available for ChatGPT Business and Enterprise/Edu customers on ChatGPT web." It adds that "Pro users can connect custom MCP servers with read/fetch permissions." Mobile is not supported. You no longer need Developer Mode.
Hosting costs for your server are yours in both cases.
Security and admin controls
Shared foundation. Both paths rely on the MCP authorization model for remote servers, OAuth with scoped tokens. Tool annotations matter in both. OpenAI's guidelines require readOnlyHint, destructiveHint and openWorldHint as explicit booleans. It adds that annotations "do not grant permission or replace authentication." Our tool annotations guide explains each hint.
Plugin review. To publish, organization owners or members with "Apps Management Write" submit a ZIP. Individual or business verification is required, and the server must pass domain verification. Plugins need a published privacy policy and must collect only the data the task needs.
Workspace controls in ChatGPT. The help article says only admins and owners can create and publish custom MCP plugins in Business workspaces. Enterprise/Edu admins can use role-based access control to choose who can connect servers and which actions apps may take.
Plain MCP. No outside reviewer checks your server. You own the threat model, from prompt injection to token handling. The MCP Apps sandbox helps, but it does not make an untrusted server safe.
Which to choose
Pick a plain MCP server if:
- You want one integration for Claude, Cursor, VS Code, ChatGPT and other clients.
- The server is internal, or you are still piloting with a few customers.
- You sell digital goods or subscriptions, which OpenAI's directory rules do not allow inside plugins.
Pick the ChatGPT Apps SDK (plugins) if:
- Your customers already work in ChatGPT or Codex and you want a directory listing.
- You want to ship skills and tools together as one installable item.
- You need ChatGPT-only UI features such as Instant Checkout for physical goods.
Do both if you can. Build a standards-first MCP server with MCP Apps UI, then package it as a plugin. OpenAI's own docs recommend that order. For the client side, see Claude vs ChatGPT for MCP.
Next step
eSEOspace builds MCP servers that work across clients and can be packaged for ChatGPT. Talk to our MCP development team, or browse more MCP comparisons. For OpenAI's wider MCP support, read OpenAI and MCP.
Frequently asked questions
Is the ChatGPT Apps SDK built on MCP?
Yes. OpenAI's docs describe a plugin's MCP server as the part that "exposes tools and connects to external systems." UI uses the open MCP Apps standard. A plugin can also be skills only, with no server.
Did OpenAI rename the Apps SDK?
OpenAI's developer docs now use "Plugins," and the old apps-sdk address redirects there. The help center still mentions the Apps SDK, and the UI library is still called @openai/apps-sdk-ui. Expect both names for a while.
Can a ChatGPT plugin's UI work in Claude?
It can if you build on the MCP Apps standard. OpenAI says this approach "lets the same UI run in ChatGPT and other compatible MCP Apps hosts." ChatGPT-only window.openai features will not work elsewhere, so feature-detect them.
Which ChatGPT plans can use custom MCP servers?
OpenAI's help center lists Business and Enterprise/Edu on ChatGPT web for full MCP support. Pro users can connect custom servers with read and fetch permissions only. Mobile is not supported as of October 2026.
