Skip to content
Kitbay
Submit

What is an MCP server?

By the Kitbay editorial team ·

An MCP server is a program that gives AI applications, such as Claude, ChatGPT, Cursor or VS Code, access to tools and data from another system through the open Model Context Protocol. It might let an assistant search your files, read a calendar, query a database or create an issue in a project tracker, with each action described in a standard way any MCP-compatible app can understand.

MCP in one paragraph

MCP (Model Context Protocol) is an open standard for connecting AI applications to outside systems. Anthropic introduced it in November 2024 to replace one-off integrations with a common method. The project's documentation compares it to a USB-C port for AI: one agreed connector, many devices. Today it is supported by a wide range of clients, including Claude Code, Claude Desktop, ChatGPT, Cursor and VS Code.

Hosts, clients and servers

MCP describes three roles:

  • Host: the AI application you use, such as Claude Desktop or VS Code.
  • Client: a connector inside the host. The host creates one client for each server it connects to.
  • Server: the program that offers capabilities to the client.

So if VS Code connects to two servers, it runs two clients, each with its own connection. The word "server" refers to the role, not to where the program runs; an MCP server may well be running on your own laptop.

What an MCP server offers

Servers provide three kinds of building blocks:

Building block What it is Who decides when it is used Example
Tools Functions that perform actions or queries The model "Search flights", "create calendar event", "send message"
Resources Read-only data given to the model as context The application A document, a database schema, a calendar
Prompts Ready-made instruction templates The user "Summarise my meetings", often shown as a slash command

Tools are the part most people notice, because they let an assistant act. A tool can write to a database, call an outside API or change files. The MCP specification says hosts must get the user's explicit consent before invoking a tool, which is why clients show approval prompts.

How a request flows

  1. You add the server to your AI app's settings.
  2. The app connects and asks the server what it offers, for example with a tools/list request. Each tool comes back with a name, a description and a schema for its inputs.
  3. You ask a question. The model reads the tool descriptions and decides one would help.
  4. The app asks your permission (depending on its settings), then sends a tools/call request with the inputs.
  5. The server does the work, often by calling the service's own API, and returns the result.
  6. The model uses the result in its answer.

Messages use the JSON-RPC 2.0 format. If you want to compare this with a direct API integration, see MCP vs API.

Local and remote servers

Local Remote
Runs On your computer On the provider's infrastructure
Transport stdio Streamable HTTP
Typical setup A command such as npx … or uvx … in a config file A URL plus a sign-in
Usually serves One client Many clients

The MCP documentation gives two examples. The file system server runs locally when Claude Desktop starts it, and works only with folders you list. The Sentry server is hosted by Sentry and reached over the internet.

What people use MCP servers for

The MCP documentation lists examples such as an assistant working with Google Calendar and Notion, Claude Code building a web app from a Figma design, and an AI model creating 3D designs in Blender. On Kitbay, the matching tool pages are Google Calendar, Notion, Figma and Blender.

Other servers are not tied to one product. They cover general jobs such as web search, browser automation, file system access and memory across conversations.

Official and community servers

Anyone can publish an MCP server. Some come from the company behind the tool; many come from independent developers. Kitbay marks a listing as official only with evidence that the tool's own vendor publishes it, and collects those on the official MCP servers page. Community servers can be just as useful, but it is worth checking who maintains them.

MCP servers and skills

An MCP server gives an agent access. An agent skill gives it a written procedure for a task. Many setups use both; our guide to MCP vs skills explains when to use which.

Safety note

A local server runs with your user account's rights, so it can do anything with your files that you can. A remote server can act through whatever account you sign in with. Install servers only from sources you have reviewed, grant the narrowest access you can, and read approval prompts before accepting them. Our MCP security questions give a step-by-step checklist.

Next steps

Questions

What does MCP server mean?
MCP stands for Model Context Protocol. An MCP server is any program that offers tools, data or prompt templates to AI applications using that protocol, whether it runs on your computer or online.
Do I need to code to use an MCP server?
No. Using a server usually means adding a short entry to your AI client's settings, such as a command or a URL. Building one does take programming.
Is an MCP server the same as an API?
No. An API is one service's own interface; an MCP server presents capabilities in a standard way that any MCP client can discover and use, and it often calls the service's API behind the scenes.
What is the difference between a local and a remote MCP server?
A local server runs on your computer, started by your AI client over stdio. A remote server runs on the provider's infrastructure and you connect to it by URL over Streamable HTTP, usually with a login.

Sources

More guides