
Ask ten people what an MCP connector is and you’ll get answers that sound different but describe the same thing. One calls it a plugin. Another calls it an integration. A third says it’s just an MCP server. The confusion matters, because the word “connector” now shows up in the settings of nearly every AI assistant your employees use, and each one is a path from an AI model into a business system.
This guide explains what MCP connectors are, how they work, how they handle sign-in, and what changes when you need to run them across a whole organization instead of one laptop.
What are MCP connectors?
An MCP connector is the link between an AI application and an MCP server. It lets the AI client discover the tools and data that server exposes, sign in to it, and call those tools on a user’s behalf through the Model Context Protocol.
The Model Context Protocol (MCP) is an open standard for connecting AI applications to external systems. Instead of building a custom integration for every pair of AI tool and business app, a system exposes its capabilities once through an MCP server, and any MCP-compatible client can use them.
“Connector” is the word many AI products use for that connection in their interface. Anthropic, for example, describes custom connectors as a way to “connect Claude directly to the tools and data sources that matter most to your workflows,” built on remote MCP.
MCP connector vs. MCP server vs. MCP client
The three terms are easy to blur. Here’s how they fit together.
| Term | What it is | Example |
|---|---|---|
| MCP server | The service that exposes tools, data, and prompts from a system | A CRM’s MCP server with tools like “find account” and “update opportunity” |
| MCP client | The AI application that calls those tools | An AI assistant, a coding tool, or a custom agent |
| MCP connector | The configured link between a client and a server, including its address and sign-in | The CRM connection a user enables in their assistant’s settings |
A useful shorthand: the server is what you connect to, the client is what does the connecting, and the connector is the connection itself.
How MCP connectors work
When a user or admin adds a connector, the same few steps happen behind the scenes.
- Configure. The client is given the server’s address, or a command to start it locally.
- Authenticate. The user signs in, usually through OAuth, or the connector uses a stored credential.
- Discover. The client asks the server which tools it offers. Each tool comes with a name, a description, and an input schema.
- Call. When a task needs a tool, the model chooses it, the client sends the request, and the server acts on the underlying system and returns a result.
The tool descriptions in step 3 matter more than most people realize. The model reads them to decide what to call, which means they act a little like instructions.
Local vs. remote connectors
MCP connectors come in two broad forms.
- Local connectors run an MCP server on the user’s own machine, typically started by the client. They’re quick to set up for a single developer, but each one lives on one device, with its own configuration and credentials.
- Remote connectors point to an MCP server hosted on the internet or inside your network. Anthropic’s help center notes that remote servers run on the internet, while earlier MCP servers ran only on a user’s computer.
Remote connectors are what make MCP practical for teams. One server can serve many users, updates happen in one place, and access can be managed centrally.
How MCP connectors handle sign-in
Authentication is where most connector decisions get made, and where most enterprise risk sits.
Per-user sign-in with OAuth
With OAuth, each user signs in to the underlying app and the connector acts with that user’s permissions. Anthropic’s guidance describes OAuth as letting the assistant act for you without seeing your password. If a user can’t see a record in the CRM, the connector can’t either.
Shared credentials
Some MCP servers use API keys or other fixed credentials. That’s simpler to configure, but it changes who the connector acts as. Anthropic notes that on Team and Enterprise plans, everyone using a connector with a shared credential shares that credential’s access. A sales intern and a sales VP would see the same data.
For anything beyond low-risk, read-only data, per-user sign-in is the safer default.
MCP connector examples
Common connectors in enterprise AI programs include:
- Productivity and collaboration: email, calendar, documents, and chat.
- CRM and customer data: account lookups, opportunity updates, and case history.
- Ticketing and ITSM: creating, triaging, and closing tickets.
- DevOps and code: repositories, pull requests, and CI pipelines.
- Data warehouses: running governed queries against analytics data.
- Finance and ERP: invoices, purchase orders, and approvals.
Each of these is useful, and each gives an AI model a way to act on systems your business depends on.
What changes when MCP connectors scale
One developer with three connectors is easy to manage. A thousand employees choosing their own connectors across several AI clients is not. Four problems tend to show up.
Connector sprawl
Every assistant, IDE, and agent framework has its own connector settings. Without a central view, nobody knows which servers are connected, who connected them, or which credentials they use.
Over-broad tools
A connector usually exposes every tool its server offers. Most teams need a few. Anthropic’s own guidance is to disable tools you don’t need, because tools can read, change, or delete data.
Untrusted servers
Anthropic warns that custom connectors can reach “arbitrary services that have not been verified,” and that malicious servers may hide instructions that push the model toward unintended actions. Because models read tool descriptions, a poisoned description can steer behavior without the user noticing.
No record of what happened
When connectors live inside individual clients, there’s no single log of which agent called which tool, with what result. That makes incident response and audits slow.
How to govern MCP connectors in the enterprise
- Approve servers before anyone connects. Keep a catalog of reviewed MCP servers, including the custom ones your teams build.
- Scope access by role, down to the tool. Give each role only the tools, and where possible the parameters, it needs.
- Prefer per-user sign-in. Use OAuth so connectors inherit each user’s existing permissions.
- Scan tool definitions. Check tool names and descriptions for hidden prompt injection before they reach a model.
- Route connectors through one gateway. A single governed endpoint gives every client the same policies and produces one log of every call.
How Airia governs MCP connectors
Airia brings MCP connectors under one governed layer through its MCP Gateway. Admins approve which servers from Airia’s catalog of pre-configured MCP servers, plus any custom servers, can be used at all, and connect the credentials those apps need.
A gateway bundles one or more approved apps behind a single endpoint for clients such as Cursor, Claude Code, and Claude Desktop, or for an agent built in Airia. With the Dynamic MCP Gateway, one address serves every user with a tool set matched to their role. Admins assign access at the app, tool, and parameter level by role, and users adjust their own preferences within those limits.
Every tool’s name and description is scanned for hidden prompt injection before it’s exposed, and tool annotations let admins filter for read-only or destructive actions. Every tool call, whether it succeeded, was denied, or errored, is logged and searchable.
Because the gateway runs on the same platform as Airia’s AI guardrails, agent constraints, and AI discovery, connector policy sits alongside the rest of your AI governance. Browse the Airia integrations library to see what’s available.
Conclusion
MCP connectors turn AI assistants and agents into tools that can act on real business systems. That’s the value, and it’s also the risk. Treat each connector as access, not just convenience: approve the servers, scope the tools, sign in per user, and route every call through a gateway you can see into. That’s how MCP connectors scale from one laptop to the whole organization.
Give every team the connectors it needs, and nothing more. See how Airia approves MCP servers, scopes tools by role, and logs every call from one gateway. Request a demo today.
Put these ideas to work.
Schedule a 30-minute walkthrough with our team.