Glossary
MCP Connector
What Claude and ChatGPT call a remote MCP server once a user has added it. A connector is a URL that speaks the Model Context Protocol plus the login that says who is calling.
General definition
An MCP Connector is the product name that AI assistants such as Claude and ChatGPT give to a remote MCP server that has been added to the assistant, together with the login that authorises it. Under the hood a connector is nothing more than a URL that speaks the Model Context Protocol over Streamable HTTP; the word connector describes how it is packaged for end users, who add it from Settings, Connectors rather than by editing a configuration file.
Two kinds exist in practice:
- Listed connectors: published by the vendor, reviewed by the assistant maker and shown in its directory. Users click to connect and sign in through the vendor’s OAuth login.
- Custom connectors: the user pastes a server URL themselves. Claude calls these custom connectors; ChatGPT exposes them through developer mode. Authentication is either OAuth (the assistant runs the authorisation flow and shows which permissions the server requests) or none, in which case the URL itself usually carries a key.
Once connected, the assistant lists the server’s tools and can call them inside a conversation. What a connector can reach is what the server’s tools expose for the identity behind the login: with OAuth that is bounded by the scopes the user approved; with a keyed URL it is whatever the key allows. Both assistants let users toggle individual tools and confirm sensitive calls. The vendor guides say to connect only to servers you trust, because a malicious server can attempt prompt injection or read whatever the model passes to it.
To remove a connector, delete it from the same connectors list; with OAuth the grant can also be revoked on the server side. ChatGPT connectors originally required two extra read-only tools, search and fetch, for deep research; they are no longer mandatory for a connector to work, but servers that want to be usable in both assistants typically still implement them.
In the Ethora ecosystem
Ethora offers both routes. In the Ethora web app under Account, AI Assistants, a user creates an API key and copies a personal connector URL of the form https://mcp.chat.ethora.com/mcp/k/<key>; pasted as a custom connector into Claude or ChatGPT, every conversation is authenticated with no login step. The key acts as the user until revoked, so the URL should be treated like a password. For listing in a directory the server also exposes an OAuth 2.1 endpoint at /mcp/oauth, where the assistant runs the login and the user approves read, write or admin scope; grants are visible and revocable under Connected AI apps.
Step-by-step pages: Ethora in Claude and Ethora in ChatGPT. The server implements search and fetch, carries titles and behaviour annotations on all 90 tools, and never offers a tool that moves money or tokens on the hosted surface, which is what the Claude and ChatGPT directory reviews require. Once connected, an assistant can create an app, a room and an AI agent and return the website widget snippet in one conversation.