What is an MCP connector
What is an MCP connector? Definition and examples
What an MCP connector is, how it differs from an MCP server and an MCP gateway, how it works step by step, and examples across CRM, ERP, HR and BI systems.
Short answer
An MCP connector is the link that lets an AI assistant use another system through the Model Context Protocol. On the assistant’s side it is an entry in the connector list; behind it is an MCP server that exposes the system’s data and actions as tools. When a user asks a question, the assistant picks a tool, the server calls the system with the user’s access, and the result comes back into the conversation.
Updated
MCP connector, defined
The Model Context Protocol (MCP) is an open standard for connecting AI applications to tools and data. It defines how a client, such as Claude, ChatGPT or an agent framework, discovers what a server offers and calls it. “Connector” is the product word assistants use for that link. Claude and ChatGPT both list connectors in their settings, and remote connectors in both are MCP servers.
So when someone says “the Salesforce connector for Claude”, they usually mean three things together: an MCP server that knows how to talk to Salesforce, the sign-in that ties it to the user’s Salesforce account, and the entry in Claude that lets the model call it.
Connector, server, client and gateway
| Part | What it is | Example |
|---|---|---|
| MCP client | The application the user talks to. It holds the conversation and decides when to call a tool. | Claude, ChatGPT, Microsoft 365 Copilot, a Gemini-based agent, an internal agent |
| MCP server | A program that exposes tools, resources and prompts over MCP and calls the underlying system. | A server that can list opportunities or run a report |
| MCP connector | The configured link between a client and a server, including the server URL and the user’s sign-in. | “Salesforce” in a user’s Claude connector list |
| MCP gateway | One managed endpoint in front of many servers that adds authentication, policy, routing and audit. | A company-wide entry point for every assistant |
Local connectors run an MCP server on the user’s own machine over stdio. Remote connectors call a server over HTTPS, which is what hosted assistants and most enterprise deployments use. See MCP gateway for the managed version.
How an MCP connector works, step by step
- An administrator approves the connector for the organization, and the user adds it in their assistant.
- The user signs in. For remote servers this is an OAuth flow in the browser, so the assistant receives a token for that user, not a shared key.
- The client asks the server which tools it offers and shows their names and descriptions to the model.
- The user asks a question. The model chooses a tool and fills in its arguments, for example an account name and a date range.
- The server checks the token, calls the system’s API with the user’s access, and returns the result.
- The model writes the answer from the result. If the tool would change data, the client asks the user to confirm first.
MCP connector examples across business systems
What a connector is worth depends on the questions it can answer. Typical ones by system:
CRM
“Which opportunities over $50,000 have not moved stage in 30 days?” from Salesforce or HubSpot.
ERP
“What is AR aging by customer, 30, 60 and 90 days?” from NetSuite, SAP S/4HANA or Oracle Fusion.
HR
“How has headcount by department moved since January?” from Workday or SAP SuccessFactors, with compensation fields kept to the people allowed to see them.
IT service
“Which priority 1 incidents breached SLA last week, by assignment group?” from ServiceNow.
BI and data
“What does the active customers metric mean, and what is it this month?” from Tableau, Power BI or Snowflake.
Knowledge
“Find the latest signed master agreement for this customer” from SharePoint or Google Drive.
What separates a good MCP connector from a demo
- User-level access. Each user signs in with their existing account, or the company’s SSO, and the system’s own permissions decide what comes back.
- Tools shaped around questions. A single “run any query” tool is flexible and hard to govern. Tools that match real questions are easier for the model to choose and for admins to review.
- Definitions. Business terms such as bookings, churn or headcount need one approved meaning, or every assistant will compute its own.
- Evidence. The result should carry its source and query, so a reviewer can check it.
- Write safety. Reads and writes separated, writes off by default, confirmation before any change.
Nexatron’s MCP connectors
Nexatron is one MCP connector that reaches 279 business systems and data platforms through a single hosted server. Users sign in with their existing account in each system, or through the company’s SSO if it has one, and Nexatron inherits their existing permissions, so every answer contains only data they’re already authorized to see. Answers use approved definitions and show the definition, source and query behind each number, and one question can combine systems. The same governed answers reach Claude, ChatGPT, Microsoft 365 Copilot and Gemini, and any MCP-capable assistant or agent. Browse the connector catalog or read how the hosted MCP server works.
Frequently asked questions
- What is an MCP connector?
- The link that lets an AI assistant use another system through the Model Context Protocol: an MCP server that exposes the system as tools, plus the user’s sign-in and the entry in the assistant’s connector list.
- Is an MCP connector the same as an MCP server?
- Closely related. The server is the program that exposes tools. The connector is the configured connection from an assistant to that server, including the user’s sign-in. Assistants such as Claude and ChatGPT call these connectors.
- Do MCP connectors work with every AI assistant?
- Any assistant or agent that supports MCP can use an MCP server. Claude, ChatGPT and Microsoft 365 Copilot all support remote MCP servers in some form, as do the Gemini API and Google’s Agent Development Kit and many other agent frameworks.
- Are MCP connectors read-only?
- Not necessarily. A server can expose tools that change data. Well-run deployments keep writes off by default and require confirmation before any change.
- How many MCP connectors does a company need?
- One per system if each system has its own server, which quickly becomes dozens. A governed layer such as Nexatron gives assistants one connector that reaches every connected system.
One MCP connector for your business systems
See Claude, ChatGPT, Copilot or Gemini answer questions from your CRM, ERP and warehouse under your users’ existing permissions.