Claude connectors vs ChatGPT plugins: one MCP server, two assistants

The short answer

Claude calls them connectors. OpenAI now calls them plugins, after earlier names of connectors and apps. Under both names, a custom integration is an MCP server: a web service that publishes tools the assistant can call. That shared foundation is the most useful fact for a business, because it means you do not have to build twice.

These are not the original ChatGPT plugins from 2023, which were retired in 2024. Today's plugins are MCP-based and managed by workspace admins.

Where they are the same

Both assistants discover the tools a server offers, read their descriptions and schemas, and call them when a question needs them. Both support remote servers over HTTPS with OAuth sign-in, and both respect read and write labeling so the assistant can ask before taking an action.

• Same protocol and the same server code
• Per-user sign-in so each person sees only their own data
• Tool annotations that mark read-only, destructive, and open-world tools
• The same need for narrow tools, bounded results, and clear descriptions

Where they differ

Claude: Anthropic documents custom connectors on every plan, including one on Free. On Team and Enterprise, an owner adds the connector in Organization settings, and members connect individually.

ChatGPT: OpenAI supports custom MCP plugins on Business, Enterprise, and Edu workspaces, subject to admin permissions. In the current Plugins experience you add one from ChatGPT Plugins as a custom MCP server, without turning on Developer Mode first. Admins choose which plugins are available and to whom, and OpenAI offers compliance logs of tool calls. Custom connectors with search and fetch tools can also feed ChatGPT's company knowledge feature.

Both vendors change these rules often. Check the current documentation for your plan before you commit to a rollout date.

Building one server for both

Design the tools once around your workflows, keep authorization in the server rather than in prompts, and then test in each client. In practice the adjustments are small: tool descriptions that one model reads more literally than the other, a search and fetch pair for ChatGPT's knowledge features, and the slightly different admin steps to publish in each workspace.

If you plan a public listing for your customers, both directories ask for explicit tool annotations, test cases, and a demo account, so it is worth building to those requirements from the start.