What the names mean
| OpenAI term | Plain-English meaning |
|---|---|
| Plugin | The installable and discoverable package. It can contain apps, skills, templates, an MCP connection, and optional UI. |
| App | The integration that connects ChatGPT or Codex to an external service, data, or actions. |
| MCP server | The protocol endpoint exposing tools and, optionally, resources or an interactive interface. |
| MCP App | An optional interactive UI—such as a form, chart, or picker—returned by an MCP tool and rendered in chat. |
| Connector | A term still used in older ChatGPT material and in the OpenAI API for OpenAI-maintained service connections. |
How a custom connection reaches ChatGPT
- Build an MCP server with focused tools and a Streamable HTTP endpoint.
- Expose it through public HTTPS, usually at a URL ending in
/mcp. - Enable developer mode where the plan and workspace policy allow it.
- Create the custom plugin/app, review the tools ChatGPT discovers, and authenticate each user to the underlying service.
- For directory distribution, complete OpenAI’s verification and review requirements.
ChatGPT does not directly launch an arbitrary local stdio server. During development, a public HTTPS tunnel can bridge to a local server. OpenAI’s Secure MCP Tunnel also supports outbound-only access to private or on-premise services without opening an inbound firewall rule; that private route is separate from public directory publishing.
Where the interface fits
Install
Plugin
What the user or admin finds and enables.
Connect
App
The relationship with an external service.
Act
MCP tools
The capabilities the model can call.
Display
MCP App UI
Optional sandboxed interface for review or interaction.
New portable UI work should use the open MCP Apps conventions. ChatGPT-specific window.openai features can add capabilities, but a well-designed tool should remain useful in clients that only render text or structured results.
Permissions and administration
- Installing a plugin, allowing its app, and authorizing the external service are separate controls.
- Workspace plans can restrict developer mode, publishing, app access, user roles, and specific actions.
- Each person normally authenticates their own account; the MCP server must validate every access token and enforce downstream permissions.
- ChatGPT permission modes may ask before any change, important actions, or other configured thresholds. Those prompts do not replace OAuth scopes or server-side authorization.
- An unreviewed custom server remains third-party code and data handling merely because ChatGPT can connect to it.
Progressive tool exposure
The OpenAI Responses API has verified Tool Search and deferred MCP loading. That lets an API application search a catalog and load only relevant definitions. OpenAI does not document the same developer control for every custom plugin inside the ChatGPT product host, so the two surfaces should not be presented as identical.
Good fit
Use ChatGPT when the workflow should live in a general business assistant, needs per-user service authorization, and benefits from the Plugin Directory or an interactive in-chat interface. Validate the exact plan, region, workspace policy, approval setting, and target surface before promising availability.