Glossary

What Are MCP Servers? A Plain Definition

Checked 2026-10-31

Loading poll...

An MCP server is a program that offers tools, data and prompts to AI applications through the Model Context Protocol (MCP), an open standard. The AI application, called the host, connects to the server with a client, asks what it offers, and calls its tools when the model needs them. One server works with many hosts. Checked 31 October 2026.

The definition in a table

Term	Meaning
MCP	Model Context Protocol, an open standard for connecting AI applications to external systems
MCP host	The AI application, such as Claude Code or a desktop assistant
MCP client	A connection object the host creates, one per server
MCP server	The program that exposes tools, resources and prompts
Tool	A function the model can call, such as "create issue"
Resource	Context or data a user or model can read
Prompt	A templated message or workflow the server offers
Transport	How messages travel: stdio for local, Streamable HTTP for remote

What it is, in plain words

The MCP site describes the protocol as "an open-source standard for connecting AI applications to external systems," and compares it to a USB-C port: one standard way to plug many things into many devices. Before it, every AI application needed its own integration for every service: one for GitHub, one for a database, one for a calendar. With MCP, someone writes a server once, and any application that speaks MCP can use it.

An MCP server is that piece. Think of it as a small adapter that sits next to a service. A GitHub server knows how to list issues and open pull requests. A filesystem server knows how to read and write files in a folder. A browser server knows how to click and type on a page. Each tells the host what it can do in a standard way, and the host passes that list to the model.

How it works

The architecture page of the MCP docs describes a client-server design. An MCP host creates one MCP client for each server, and each client keeps its own connection. The data layer uses JSON-RPC 2.0. The steps in a normal session are:

Connect. The host starts or reaches the server.
Discover. The client asks the server for its tools, resources and prompts.
Present. The host gives the model the tool names and descriptions.
Call. When the model decides to use a tool, the host sends the call to the server and may ask the user to approve it.
Return. The server runs the action and sends back a result, which goes into the model's context.

The server may also ask the user for more information, which the spec calls elicitation.

The building blocks

The spec lists these primitives:

Tools. Functions the model can call. This is the part people mean most often. Examples: search issues, run a query, take a screenshot.
Resources. Data the host can read as context, such as a file, a record or a schema.
Prompts. Reusable templates a user can pick, such as "review this pull request."
Elicitation. A request from the server for the user to supply something it needs.

Local and remote servers

MCP servers talk over two transports.

	Local (stdio)	Remote (Streamable HTTP)
Where it runs	A process on your machine	The service's own infrastructure
Network	None	HTTPS
Typical auth	Environment variables, local keys	OAuth, bearer tokens, API keys, custom headers
Good for	Files, local tools, scripts	Hosted services like project trackers and docs
Setup	A command the host starts	A URL the host connects to

Claude Code shows both forms. A remote server is added with a URL:

claude mcp add --transport http notion https://mcp.notion.com/mcp

A local stdio server is added with a command after a double dash:

claude mcp add --env AIRTABLE_API_KEY=YOUR_KEY --transport stdio airtable -- npx -y airtable-mcp-server

The -- separates Claude's own options from the command that runs the server. Servers can also be defined in a .mcp.json file at a project root. Some services still expose only an older SSE endpoint, and Claude Code can connect to those too.

What MCP servers are good for

The Claude Code docs give examples of what a connected server enables: querying a database, creating drafts in an email account, working with issue trackers and monitoring tools. On this site we group servers by job:

Job	Examples on this site
Code hosting and review	GitHub MCP
Browser testing	Playwright MCP
Docs lookup	Context7
Files	Filesystem MCP
Project work	Linear, Notion, Atlassian Jira
Errors and monitoring	Sentry
Web search and scraping	Exa, Firecrawl

The full list is at MCP servers, with a ranked view at best MCP servers.

What an MCP server is not

Not a model. It does not think. The model decides when to call it.
Not an agent. It exposes capabilities; an agent chooses how to use them. For how agents talk to each other, see A2A protocol.
Not the same as an API. An API is a service's own interface for developers. An MCP server wraps it for AI applications. See MCP vs API.
Not a skill. A skill is a folder of instructions. A server provides live tools. See agent skills and skills vs MCP.
Not a plugin store. Servers vary in quality, and a directory is not a safety review.

Cost: tokens, not subscriptions

Most servers are free to run, and some hosted ones carry a service plan. The hidden cost is context. Each connected tool needs a name and a description in the model's context window. Claude Code defers MCP tool definitions by default, so only tool names and server instructions enter context until a tool is used. Its docs also say command-line tools such as gh and aws can be more context-efficient than MCP servers. If your agent already has a good CLI for a job, you may not need a server. Our context window entry explains why this matters.

Risks and safe habits

The spec is blunt that tools represent arbitrary code execution, and that hosts must get user consent before invoking them. Claude Code's docs add: verify you trust each server before connecting it. Servers that fetch external content can expose you to prompt injection, where text on a web page or in a ticket tries to steer the model.

Habits that help:

Install servers from the vendor or a source you can inspect.
Give each server the narrowest token or scope it needs. Read-only first.
Keep write tools behind approval.
Do not connect a server that reads untrusted content to an agent that can also send data out, without limits.
Review project-level .mcp.json files in cloned repositories before approving them. Claude Code requires you to trust the workspace first.
Remove servers you no longer use.

How to judge a server before you install it

Question	Why it matters
Who publishes it?	A vendor's own server is accountable; an anonymous one is a stranger's code
What licence and source?	You can read what it does if the source is open
Local or remote?	Local runs code on your machine; remote sends your requests to someone else
What auth does it need?	OAuth scopes or a token you can limit are better than a master key
Which tools does it expose?	Fewer, read-only tools are easier to trust
Can it write or delete?	Write tools need approval and a backup plan
What does it cost in context?	Many tools mean many descriptions in the context window
When was it last updated?	A server for a changing API goes stale fast

Our add-on pages follow this list: each records the publisher, licence, install lines for several agents, tools, auth and risks, with a check date.

A first session, step by step

Suppose you add a GitHub server and ask, "Which pull requests have been open longest?" The host starts the client, the client lists the server's tools, and the model sees names like "list pull requests." It picks one, the host asks you to approve it if your settings require that, and the server calls GitHub with your token. The result comes back as text, and the model answers. Nothing special happened to the model. It read a tool description, produced a structured call, and read the reply. That is the whole idea, and it is why the quality of a tool's name and description matters as much as its code.

Who looks after the standard

MCP is an open standard with a versioned spec. The version listed on the MCP site when we checked is 2026-07-28. The A2A project's August 2026 announcement says MCP, goose, AGENTS.md and A2A all sit under the Agentic AI Foundation, directed by the Linux Foundation. That neutral home is one reason many vendors support it.

How to get started

Pick one job you repeat, such as reading pull requests.
Choose a server that does it, from MCP servers.
Follow the install steps in install an MCP server in Claude Code.
Start with read-only access and test on a non-critical project.
If you need your own, see how to build an MCP server.

Our take

An MCP server is a small, useful idea: write the connection once, use it everywhere. Its risk is the same as its strength, since it gives a model hands. Treat each one like software you install, with a scope, a source you trust and a reason to exist. We have documented servers from their vendors' pages and have not tested them ourselves. See also MCP vs RAG and MCP vs A2A.

Questions

What is an MCP server in one sentence?

A program that gives an AI application access to tools, data or prompts through the Model Context Protocol, so the application does not need custom code for each service.

Is an MCP server the same as an API?

No. An MCP server usually sits in front of a service's API and presents it in a standard shape that AI applications can discover and call.

Are MCP servers safe?

Only as safe as you make them. The spec says tools represent arbitrary code execution, and Claude Code's docs say to verify you trust each server, especially ones that fetch external content.

Do I need to run MCP servers on my own machine?

Not always. Local servers use stdio on your machine, while remote servers use Streamable HTTP and are hosted by the service, for example mcp.notion.com.