Beginner's guide

What is an MCP server? A beginner's guide

MCP gives AI applications a common way to connect to tools and data. This guide explains the connection, what to check before installing a server, and how to test one without granting more access than it needs.

Instructions checked: July 23, 2026

MCP in plain English

Model Context Protocol, or MCP, is an open protocol for connecting an AI application to outside capabilities. The application might use an MCP server to read a file, search the web, query a database, or call a development tool.

The server is the connector. Installing one does not make the underlying service free, trusted, or safe by default. It tells a compatible AI client how to reach a particular set of tools or context.

The three parts of an MCP connection

01

Host

The AI application you use, such as a coding editor or chat client.

02

Client

The part of the host that manages one connection to an MCP server.

03

Server

The program that exposes tools, resources, or prompts through MCP.

A server can expose tools that perform actions, resources that provide context, and prompts that offer reusable interaction templates.

Local and remote servers

Local server

Your client starts a command on your computer. That process may inherit access to local files, environment variables, or credentials.

Remote server

Your client connects to a hosted URL. You may need to sign in or provide a token, and the remote operator may process the data you send.

MCP server, Agent Skill, plugin, or API?

TypeWhat it providesWhat to verify
MCP serverA live connection to tools or dataPermissions, operator, code, and transport
Agent SkillInstructions and reusable workflow knowledgeSource files and actions the instructions request
PluginA packaged extension for a specific applicationPublisher, requested scope, and update path
APIA service interface used by softwareAuthentication, cost, rate limits, and data policy
Before you connect

A safer six-step setup

  1. 01

    Confirm the source

    Use the maintainer's repository or official service documentation. Check that the package name and URL match.

  2. 02

    Read the requested access

    List the files, accounts, network services, and actions the server can reach.

  3. 03

    Check cost and data handling

    A free server package may call a paid API. Find out where prompts, files, and results are processed.

  4. 04

    Use limited credentials

    Start with a test account, a narrow token, or a restricted folder when the service supports it.

  5. 05

    Add it through your client

    Follow the current instructions for your AI application. Local commands and remote URLs use different settings.

  6. 06

    Test and keep a removal path

    Try a harmless request first. Record how to remove the configuration and revoke its credentials.

Official client references

There is no universal install command. Use the documentation for the client and server you chose. Product support and plan availability can change.

Test, troubleshoot, and remove

After setup, confirm that the client can list the server and its tools. Start with a read-only or disposable task. If the connection fails, check the command or URL, authentication, client logs, and the server's supported transport.

To stop using a server, remove its entry from the client, stop any local process, and revoke tokens or OAuth access at the connected service. Removing the client entry alone may leave a credential active.

Common questions

MCP FAQ

What is an MCP server?

An MCP server is a program that makes tools or context available to an AI application through the Model Context Protocol. It can connect the application to services such as files, databases, browsers, or developer tools.

How does MCP work?

An AI application acts as the host. It creates an MCP client that connects to a server, discovers what that server provides, and sends requests when you approve or trigger an action. The server returns results in a standard format.

How do I install an MCP server?

Installation depends on both the server and your AI application. Some servers run as a local command, while others use a remote URL. Use the server's official instructions, add it through your client's MCP settings, and test it with the least access it needs.

Which AI applications support MCP?

MCP support is available in clients such as Claude, Cursor, Visual Studio Code, Codex, and some ChatGPT plans or workspaces. Features and setup steps differ, so check the current documentation for your client.

Are MCP servers safe?

Not automatically. An MCP server may read data, call external services, or perform actions with your credentials. Review its source and permissions, prefer a limited test account, and remove both the configuration and any tokens when you stop using it.

What does MCP100 Selected mean?

Selected entries are editorial recommendations that MCP100 confidently recommends today based on the current editorial record. Selection is not a guarantee of safety, maintenance, compatibility, completeness, or suitability for every use.

What does MCP100 Watchlist mean?

Watchlist entries are promising candidates under consideration or requiring more evidence. They are not recommendations or rejections, and their unresolved evidence gaps remain visible.

How does MCP100 decide what appears first?

Admission determines eligibility and lifecycle. Coverage preserves useful diversity. Ranking determines presentation order. Recommendation Score is an internal Ranking signal used only after Admission and Coverage; it is not a public safety, trust, or objective quality score. Unknown evidence remains Unknown.

Choose your next MCP with evidence

MCP100 separates discovery signals from reviewed evidence. A candidate listing is not a safety certificate or an endorsement.