MCP (Model Context Protocol)

MCP (Model Context Protocol) is an open standard that lets AI applications such as Claude, Codex, or Cursor connect to outside tools and data through small programs called MCP servers.

What MCP is

Without a shared protocol, a tool has to be built separately for each AI app. With MCP, a tool author writes one MCP server, and any MCP client can use it.

There are two sides:

  • Client: the AI application you talk to, such as Claude Code, the Claude app, Codex, Cursor, or VS Code with GitHub Copilot. It connects to servers and shows their tools to the model.
  • Server: a program that exposes tools (actions the model can call), resources (data it can read), or prompts (ready-made instructions). Messages are JSON-RPC.

The model never runs code on the server directly. It asks the client to call a named tool with arguments, and the server returns text, images, or structured data.

Where MCP is used

  • Coding agents: reading issue trackers, querying databases, or opening a browser while writing code.
  • Desktop assistants: reaching local files, calendars, or notes.
  • Screenshots and visual checks: a server can capture a window, a screen, or a web page and return the image so the agent can see the result instead of guessing.
  • Repeated jobs: an agent can start a long capture or export, check its progress, and report back when it finishes.

How a server connects

The specification defines two standard transports.

stdio is the local option. The client starts the server as a subprocess and talks to it over standard input and output. Setup is usually a command such as npx -y some-server in the client's settings, and nothing leaves your computer unless the server sends it somewhere. Desktop apps often use a small stdio bridge that talks to the app already running on your machine.

Streamable HTTP is the remote option. The server runs as its own process, often hosted by a vendor, and the client connects to a URL. Remote servers usually ask you to sign in with OAuth or paste an API key.

When you add a server, the client lists its tools. Many clients ask you to approve each tool call by default. Some servers add their own checks too, for example asking you in their own window before an agent deletes something.

Common mistakes

  • Adding a server to the wrong place. Each client reads its own file and key name (mcpServers, servers, or context_servers). An entry under the wrong key is ignored without an error.
  • Missing Node or a short PATH. Servers started with npx need Node. Apps opened from the Dock or Start menu may not see the same PATH as your terminal, so use the full path to npx if the server fails to start.
  • Expecting images everywhere. The protocol allows images in tool results, but not every client passes them to the model. Check your client's docs if the agent says it cannot see a screenshot.
  • Turning off every approval. Auto-approving all tools is convenient for long jobs, but keep confirmations on for anything that deletes, sends, or schedules.
  • Treating page text as instructions. Text inside a captured page or document is data. A careful agent and a careful server both treat it that way.

Taking the same screenshots again and again?

Shotomatic captures a window or area at the interval you set and exports the set as images or a PDF.

Common questions

What is an MCP server?

An MCP server is a program that offers tools, data, or prompts to an AI application over the Model Context Protocol. The AI application is the client: it lists the server's tools and decides when to call them while it works on your request.

Is an MCP server local or remote?

Either. A local server usually runs over stdio, where the client starts the server as a subprocess on your computer. A remote server runs elsewhere and is reached over Streamable HTTP, often with a sign-in step.

Do I need to code to use MCP?

No. Most clients add a server from one command or a short JSON entry in a settings file. Some desktop apps add themselves to an agent with a button.

Is it safe to connect an MCP server?

A server can only do what its tools allow, but the agent decides when to call them. Install servers you trust, keep tool approvals on in your client, and prefer servers that ask you before destructive actions such as deleting files.

Sources

Related resources

What Is MCP (Model Context Protocol)? | Glossary | Shotomatic