What is an MCP server?
An MCP server is a program that makes selected tools, resources, or prompt templates available to an AI application through the Model Context Protocol. It sits between the application and an external system. The server defines what can be discovered or invoked; the client controls the connection and presents those capabilities to the model.
What job does an MCP server do?
It provides a defined interface between an AI client and another system. That system might be a local folder, a database, a web service, or a developer tool. The useful question is not whether a server adds intelligence, but which capabilities it exposes and under what permissions.
Where does it fit in the workflow?
The user configures an MCP-capable client, the client connects to the server, and the server describes its available primitives. When a tool is selected, the request travels through that connection. Approval controls, credentials, logging, and isolation remain implementation choices that should be reviewed separately.
MCP roles at a glance
| Part | Primary role | Review question |
|---|---|---|
| Host or client | Manages the user session and connections | What approval controls exist? |
| MCP server | Exposes a defined set of capabilities | What can it read or change? |
| External system | Stores data or performs the underlying action | Which credentials and scopes are used? |
Before you connect a server
- Identify every tool and data source the server exposes.
- Use the narrowest practical credentials and permissions.
- Check how actions are approved, logged, and revoked.
- Test with non-sensitive data before wider use.
Frequently asked questions
Is an MCP server the same as an AI model?
No. It exposes capabilities to a compatible application; it does not have to contain or run a language model.
Does every coding agent support MCP?
Support and configuration vary by client and version. Check the client documentation before choosing a server.
Can an MCP server run locally?
Yes, some servers run as local processes, while others are reached over a network transport. The deployment model affects trust and credential handling.
Is an MCP server automatically safe?
No. Review its code or publisher, requested permissions, data paths, dependencies, and approval behavior before use.
How should I choose an MCP server?
Start with the task, then compare compatibility, scope, maintenance signals, setup cost, and the consequences of a mistaken action.