Agents Lookup

Learn / MCP

What MCP is and how to connect your tools

The Model Context Protocol explained in plain terms: hosts, clients and servers, local vs remote, and the steps to connect your first tool.

Updated 2026-11-19. "(source N)" refers to the numbered list at the end.

The one-sentence version

MCP, the Model Context Protocol, is an open standard that lets AI apps connect to outside data and tools in one common way (source 1). The project's own comparison is to a USB-C port: one plug shape, many devices.

Before MCP, every AI app needed custom code for every service it talked to. With MCP, a service builds one server, and any app that speaks MCP can use it. Claude, ChatGPT, VS Code and Cursor are among the apps that support it (source 1).

Who runs it

Anthropic created MCP. In December 2025 it was given to the new Agentic AI Foundation under the Linux Foundation, alongside OpenAI's AGENTS.md and Block's goose agent (source 5). That move matters for buyers: the standard is not owned by one vendor.

The current version of the spec is dated 28 July 2026 (source 3).

The three pieces

MCP uses three roles (source 2).

What a server can offer

A server can give the AI three kinds of things (source 2).

Asking the user

Going the other way, a server can ask the user for more input or a confirmation through a feature called elicitation (source 2). Some older features, such as sampling and roots, were deprecated in the July 2026 version (source 2), so do not build new work on them.

Local and remote servers

A local server runs on your own computer and talks to the app through standard input and output, known as stdio. The filesystem server is a common example (source 2).

A remote server runs on someone else's infrastructure and talks over the web using Streamable HTTP. Remote servers can use normal web sign-in, and the spec recommends OAuth for getting access tokens (source 2). Anthropic's guidance for builders also points remote servers at Streamable HTTP and notes the older SSE transport is being phased out (source 6).

For a small business, remote servers run by the vendor you already use are usually the easiest and safest. You sign in with your normal account, and there is nothing to install.

How to connect a tool

In chat apps like Claude, a remote MCP server shows up as a connector. You can pick one from a reviewed directory, or add a custom one by pasting its URL (source 6). OpenAI and VS Code also document how to connect MCP servers to their tools (source 7, source 8).

In Claude Code, the command line coding agent, you add servers with one command (source 4).

Who can see a server

Claude Code also lets you choose who sees a server (source 4). Local scope is just you in one project. Project scope writes the server to a .mcp.json file that your team shares through version control. User scope makes it available to you in every project.

A first project

Pick one tool your team uses every day and that has an official MCP server, such as your issue tracker or your docs. Connect it with read-only access first. Ask the agent questions you would normally look up by hand. Once that works well, consider giving it write access for one narrow action.

Keep secrets out of shared files. Claude Code supports environment variable references in .mcp.json so that API keys are not committed to your repository (source 4).

What MCP does not do

MCP is plumbing. It does not decide how the model uses the tools, and it cannot enforce safety on its own. The spec says this directly and asks apps to get clear user consent before calling tools or sharing data (source 3). Treat each server like a new piece of software with access to your accounts. Our next article covers how to vet one.

Sources

Written by Agents Lookup in our own words. Facts come from these pages: