A ChatGPT plugin is an installable package of skills, MCP connections, or both. Skills add reusable workflows and files; MCP provides tools and data. A plugin can work without MCP. Missing MCP features must be checked against the specific ChatGPT client.
Plugins package workflows and connections
Reviewed on 2026-10-01, Asia/Bangkok.
OpenAI describes current ChatGPT and Codex plugins as installable packages. Their simplest forms are skills only, an MCP server only, or skills combined with an MCP server.
A skill contains a SKILL.md file and can include references, templates, assets, and helper scripts. It teaches the model when and how to complete a workflow. The MCP server supplies the tools, schemas, service connections, and controlled actions used by that workflow.
For example, a document workflow can use packaged instructions to process a file provided in the conversation. It needs an MCP connection only if it must retrieve information or act through an external service.
The comparison therefore has two questions:
- What does the plugin package add around a connection?
- Which MCP features does the relevant ChatGPT client actually implement?
A feature missing from a client is not automatically a limitation of plugin packaging. Equally, packaging an MCP server does not establish that the client supports every feature of the protocol.
This is the initial architectural distinction, not a completed MCP compatibility matrix. We still need to verify individual capabilities and their conditions.
Sources: