What does an MCP server do?

It advertises a list of tools (functions) and, optionally, resources and prompts. When the AI assistant decides to use a tool, it sends a structured request to the MCP server, which runs the tool and returns the result.

What is an MCP server in AI?

In AI, an MCP server is how large language models like Claude and ChatGPT reach outside data and take real actions — reading files, searching notes, querying databases, calling APIs — through a shared protocol instead of one-off integrations.

Back to home

Guide

What is an MCP server?

Q: What is an MCP server?

An MCP server is a service that exposes tools and data to AI assistants over the Model Context Protocol (MCP), an open standard introduced by Anthropic. The AI assistant is the MCP client; the server hosts the tools the assistant can call.

The short version, then the longer one

An MCP server is a service that exposes tools and data to AI assistants — Claude, ChatGPT, Codex, Cursor, and others — over the Model Context Protocol (MCP). MCP is an open standard, introduced by Anthropic in late 2024, that standardizes how models talk to the outside world.

The assistant is the MCP client. The server hosts the tools (functions the assistant can call), optionally resources (documents the assistant can read), and prompts (reusable prompt templates). When the model decides it needs to search your notes or write a file, it sends a structured tool call to the server, which runs it and returns a result.

Why MCP exists

Before MCP, every integration between an AI assistant and an outside system was a bespoke plugin. That meant every tool had to be re-implemented per client, and users couldn't move their integrations between assistants. MCP is a shared protocol — write an MCP server once, and any MCP-compatible client can use it.

What an MCP server actually does

Advertises tools. It returns a JSON list of available tools with names, descriptions, and input schemas.
Runs tool calls. When the assistant sends tools/call, the server executes the function and returns the result.
Authorizes requests. A remote MCP server typically uses OAuth so requests act as the signed-in user.
Streams updates. The transport (usually Streamable HTTP) lets the server push progress or partial results.

Local vs remote MCP servers

Local MCP servers run on your machine, launched by the client from a config file such as claude_desktop_config.json. They're great for filesystem tools and developer utilities.

Remote MCP servers run in the cloud and speak HTTPS. Clients add them through a Connectors UI and authorize with OAuth. Muninn is a remote MCP server — no config file, no local proxy.

A concrete example

Muninn's MCP server advertises tools like search_notes, create_note, and list_collections. Ask Claude "find my notes about MCP", and Claude picks search_notes, sends the query, and shows you the results. No copy-pasting, no separate app to open.

Try an MCP server in five minutes

The fastest way to see MCP work is to connect a hosted server to Claude or ChatGPT. Muninn ships one at your workspace URL — see Connect Muninn to Claude and ChatGPT for the walkthrough.