What is an MCP server?

Short answer

An MCP server is a program that exposes capabilities to AI applications through the Model Context Protocol (MCP): tools the model can call, resources it can read and prompt templates a user can pick. The AI app (the host) connects to it through an MCP client. A server runs locally on your machine over stdio, or remotely over HTTP, and any MCP-capable app can use it.

Host, client and server: three roles, often confused

MCP has three participants, and most confusion about “MCP servers” comes from mixing them up.

Role What it is Example
Host The AI application you work in. It manages the connections and talks to the model. Claude Desktop, Claude Code, Visual Studio Code
Client A component inside the host that holds one connection to one server. One client per connected server
Server The program that offers context and actions to the client. The Sentry server, the filesystem server, the GitHub server

The MCP documentation gives a concrete case: when Visual Studio Code connects to the Sentry MCP server, it creates an MCP client object for that connection. When it also connects to the local filesystem server, it creates a second client. The host then merges the tools of all its servers into one registry that the model can choose from.

So “MCP server” says nothing about size or location. It is simply the side of the connection that provides something.

What a server can offer: tools, resources and prompts

A server exposes three building blocks. Each one is controlled by someone else, and that matters for how you design and trust a server.

Building block What it does Who decides when it is used Protocol methods
Tools Functions the model can call: search, write a record, send a message The model tools/list, tools/call
Resources Read-only data for context: file contents, a database schema, documentation The application resources/list, resources/read
Prompts Reusable instruction templates, often shown as slash commands The user prompts/list, prompts/get

Every tool comes with a name, a description and a JSON Schema for its inputs. That description is what the model reads when it decides whether to call the tool, which is why a vague tool description leads to a model that never uses it.

The documentation’s example of all three together is a database server: tools to query the database, a resource with its schema, and a prompt with worked examples of how to query it.

How a call actually flows

  1. The client asks the server what it supports. Under the current specification (version 2026-07-28) every request carries the protocol version and capabilities, and a client can call server/discover to get the server’s identity and features in one go.
  2. The client sends tools/list and gets back the available tools with their schemas.
  3. The host hands those tools to the model alongside your question.
  4. The model decides to call one, for example weather_current with a location.
  5. The host routes that call to the right server with tools/call.
  6. The server does the work, usually by calling an API or database, and returns text, images or structured data.
  7. The host puts that result into the model’s context, and the model writes its answer.

The model never talks to the server directly. The host sits in between, which is where approval prompts and permission settings live.

Local and remote servers

MCP defines two standard transports:

  • stdio: the host starts the server as a subprocess on your machine and talks to it over standard input and output. This is a “local” server, such as the reference filesystem server. It typically serves one client.
  • Streamable HTTP: the server runs as its own process, often at a provider, and takes each message as an HTTP POST to one endpoint, such as https://example.com/mcp. This is a “remote” server, like Sentry’s. It typically serves many clients, and the specification recommends OAuth for sign-in.

The difference has big consequences for teams. We cover it in local vs remote MCP servers.

Examples of MCP servers

  • Reference servers from the MCP project itself: Everything, Fetch, Filesystem, Git, Memory, Sequential Thinking and Time. The project calls them educational examples, not production-ready software.
  • Vendor servers: GitHub maintains its own official server, hosted at https://api.githubcopilot.com/mcp/ or run locally in Docker, with a --read-only flag that switches off write tools. Sentry runs a remote server on its own platform.
  • Your own server for an internal system, written with one of the official SDKs. See how to build a custom MCP server.

The common misunderstanding: “MCP server” is not “an API with a new name”

Many people read MCP server as a rebranded API. The difference is who it is written for. An API is written for a developer who reads documentation and writes code against it. An MCP server is written for a model at runtime: it lists what it can do in plain language, with schemas, so any MCP host can discover and use it without custom integration code per AI tool. One server for your CRM works in every MCP-capable app your team uses, as far as that app supports the transport and sign-in method.

The other half of the misunderstanding: an MCP server does not make the model smarter or give it memory by itself. It only adds tools and context, and the model has to choose to use them. Where to find servers you can trust is covered in where to find trustworthy MCP servers.

What this means for a team

Each person connects servers in their own host. A server one colleague added to Claude Desktop on their laptop does not exist for the colleague in ChatGPT, or for the developer in Cursor. A remote server with per-person sign-in is the only kind that several people and several AI tools can reach in the same way. Even then, each connection only sees what that server exposes; MCP itself does not share what one person’s AI learned with anyone else.

Where Bombos fits

Bombos is a plugin that gives every AI tool your team uses (Claude, Claude Code, ChatGPT, Copilot, Grok, Codex, Cursor, your own agents) one shared memory. It holds team knowledge, rules the AI learned about how you work, and what already sits in mail, documents and repos, and it shows what each answer rests on. It runs in our cloud or on your own server, and one person can try it first with their own agents.

Related questions

What does MCP stand for?

Model Context Protocol. It is an open standard for how AI applications exchange context with outside systems, built on JSON-RPC 2.0.

Is an MCP server the same as an API?

No. An MCP server usually sits in front of an API and turns it into tools with names, descriptions and input schemas that any MCP client can discover and call the same way. The API stays underneath.

Does an MCP server run the AI model?

No. The model runs in the AI application or at the model provider. The MCP server only offers tools, data and prompts; the host decides what goes into the model's context.

Can one AI app use several MCP servers at once?

Yes. The host creates one MCP client per server and combines the tools of all connected servers into one list the model can choose from.

Do I need to code to use an MCP server?

Not to use an existing one: you add its URL or start command in your AI app. Building your own server for an internal system is development work.

Sources

  1. Architecture overview (Model Context Protocol)
  2. Understanding MCP servers (Model Context Protocol)
  3. Transports, specification 2026-07-28
  4. Model Context Protocol servers (reference implementations)
  5. GitHub MCP Server
The first half hour · free

What does your AI forget?

  1. 01You tell us which AI you use and where it keeps starting over.
  2. 02We show what a shared memory changes there, on your own examples.
  3. 03You keep the plan, with or without Bombos.
Piet Baudoin
Piet BaudoinYou talk to Piet, from the Bombos teamDutch team from Utrecht and Nijmegen · data in Europe
We reply the same working day.