An MCP connector is the link between an AI application and an outside system, made through the Model Context Protocol. When you add a connector to Claude, ChatGPT, or Cursor, you're pointing the AI client at an MCP server that offers tools, such as "search tickets", "get invoice", or "create task", and the AI can then call those tools during a conversation. Connectors are how AI assistants move from answering questions to working with your actual data.
This guide explains how MCP connectors work, where they come from, how to choose between vendor, community, and self-built connectors, and what to check before connecting AI to enterprise systems. For background, see what MCP is.
How MCP connectors work
Every connection has three parts:
- The MCP client: the AI application, such as Claude Desktop, ChatGPT, Cursor, or an agent framework.
- The MCP server: a service that describes its tools (names, inputs, and what they do) and runs them when called.
- The backend system: the CRM, ticketing tool, database, or API the tools act on.
When a user asks a question, the AI decides which tool to call, the client sends the request to the server, the server calls the backend, and the result returns to the AI. Remote MCP servers typically authenticate clients with OAuth, so each call is tied to an identity with defined permissions. See MCP authorization and authentication.
Where MCP connectors come from
| Source | Examples | Strengths | Watch out for |
|---|---|---|---|
| Vendor-published servers | SaaS vendors offering an official MCP server for their product | Maintained by the vendor; matches their API | Tools reflect the vendor's choices, not your processes |
| Community servers | Open-source servers on GitHub for popular tools and databases | Fast to try; wide coverage | Quality and security vary; review the code and permissions |
| Self-built servers | MCP servers on top of your own internal and partner APIs | Exactly your tools, your rules, your audit | You own building, hosting, and governing them |
| Platform-generated servers | Servers generated from APIs already published in an API management or integration platform | Reuses governed APIs; less code | Depends on the platform's governance features |
Choosing connectors for enterprise systems
Ask these questions before connecting an AI client to a business system:
- Who maintains it? Vendor, community, or your team, and how quickly are security fixes released?
- What can it do? List every tool and whether it reads, writes, or deletes.
- How does it authenticate? Prefer OAuth with scoped permissions over shared API keys.
- Whose identity does it act as? The user's, or a shared service account with broad access?
- Where does data go? Which data leaves your systems, and is it logged or retained?
- Can you see what it did? Every call should be logged with the user, tool, inputs, and result.
Building your own MCP connectors
For internal and line-of-business systems, building your own MCP server is often the right answer, because you decide exactly which tools exist and what they can touch. Good practice:
- Start from your existing APIs and their contracts, rather than writing new logic.
- Design tools around tasks, such as "get order status", rather than exposing every endpoint.
- Write clear tool descriptions, because the AI decides what to call based on them.
- Label tools by risk (read, write, destructive) and require confirmation for writes.
- Pass the caller's identity and claims to the backend, so normal access rules apply.
- Log every call and review usage regularly.
Our walkthrough on exposing an API as a governed MCP server covers this end to end.
Frequently asked questions
What is an MCP connector?
It's a connection between an AI application and an MCP server, which exposes tools the AI can call to read data or take actions in another system.
Are MCP connectors and MCP servers the same thing?
They're closely related. The MCP server is the service that offers the tools; "connector" usually refers to that server as it's added to an AI client like Claude or ChatGPT.
Are community MCP connectors safe to use?
Some are, but quality varies. Review the code and tools, run them with the narrowest permissions possible, and prefer servers that support OAuth and per-user identity.
Can I build MCP connectors for internal systems?
Yes. Many organisations build MCP servers on top of their existing internal APIs, so AI assistants can work with internal data under the same access rules as other applications.
How do I control what an MCP connector can do?
Limit the tools it exposes, give it narrow scopes, label tools by risk, require confirmation for writes, and log every call.
Connect AI to your systems on your terms
Vendor and community connectors are a fast start; governed connectors on your own APIs are what make AI safe for enterprise systems. To see Koodisi turn your APIs into governed MCP connectors, book a demo.