MCP connector: definition and examples

An MCP connector links an AI application to external tools or data through the Model Context Protocol. Products use the term for a client connection or a ready-made service integration, so it is not always a strict synonym for “MCP server.” The connection lets an agent discover available operations and receive their results.

What is an MCP connector

“Connector” is a product term rather than a separate formal network role. The MCP architecture distinguishes the host application, its client connections and servers providing capabilities. A product can call its server registration screen a connector screen, while another calls a packaged service adapter a connector.

For example, Claude can register a remote connector that points to an AIPx agent URL. Behind that endpoint, AIPx can expose permitted Notion and Gmail tools from separately connected accounts. That is one client-facing connection over more than one service integration.

The official architecture specification defines the host, client and server roles. Use those roles when reasoning about protocol behavior, and read a product’s documentation to understand its connector label.

MCP connector vs MCP server

Term	Usual meaning	Practical example
Connector	Product label for a connection or packaged adapter	A saved remote connection in an AI app
MCP server	Software implementing MCP capabilities	An authenticated tool endpoint
MCP client	The component communicating with a server	The connection inside a coding agent
Integration	The broader service connection and mapping	A connected Gmail account and its tools

A URL and a local launch command can both identify ways to reach a server. They have different deployment and authentication requirements. A client connection cannot reach private local files merely because the same application also has a desktop interface.

MCP integration

MCP integration is the work of connecting a host to server capabilities, and often connecting that server to an application API. It includes configuration, authentication, tool descriptions and handling results. Merely naming an API operation in a prompt does not establish that connection.

A useful integration also preserves the account boundary. In AIPx, rules combine the tool, connected account and agent. Hide is enforced server-side. Allow and Confirm are execution hints; Confirm requests review whose actual behavior depends on the client. Read access still deserves a confidentiality check.

See the supported integration catalog for concrete services and the MCP gateway guide for how accounts can be shared across agent connections.

MCP connectors in Claude and ChatGPT

Claude uses a connector interface for remote server connections. ChatGPT’s current setup documentation uses Plugins for custom MCP servers, although “connector” remains a common search term. Neither label changes the server’s underlying protocol role.

With AIPx, connect the service accounts, create an agent and copy its OAuth-protected URL. Add that URL through the client’s supported interface, then verify one known read. Follow the Claude setup or ChatGPT setup for current menus and account considerations.

A direct official server can be simpler when you need one service. A gateway adds another operator to trust, but can provide shared account connections and different access views for separate agents.

LangChain and LlamaIndex MCP integration

Frameworks can adapt MCP tools into their own agent tool interfaces. LangChain documents MultiServerMCPClient in its MCP adapter guidance. LlamaIndex documents MCP tool wrappers in its MCP example.

An adapter does not grant account authorization or make all operations appropriate. Configure the server’s transport and credentials, review the imported tool set and preserve the task’s access boundary. For the server’s responsibilities, read what an MCP server is. Connect free.