What is MCP (Model Context Protocol)?

How it works

MCP has two sides. An MCP server is a program that offers tools (actions such as “search” or “read transcript”) and sometimes resources (data the assistant can read). An MCP client is the AI app, such as Claude Desktop, Claude Code or Cursor, that connects to servers and lets the model call their tools.

When you ask the assistant something, the model sees the list of available tools and their descriptions. If a tool would help, it calls it with some arguments, the server runs it and returns a result, and the model uses that result in its reply. Local servers commonly talk to the client over standard input and output; remote servers use HTTP.

An example

You connect an assistant to a local search server for your recordings and ask, “Summarize what the client said about pricing in last week's calls.” The assistant calls the server's search tool with “pricing,” gets back matching transcript passages with timestamps, and writes a summary with references. The recordings themselves never moved; only the passages the tool returned were shared with the assistant.

Before MCP, every assistant had its own plugin or integration system, so a tool maker had to build a separate connector for each one. MCP replaces that with one protocol: a single server works with any client that supports it. That is why the same local server can be used from a chat app, a coding assistant and an editor.

Privacy and safety considerations

MCP makes it easy to give an assistant access to things, so it is worth thinking about what you are giving:

• Results travel to the assistant. Whatever a tool returns goes to the AI model. If that model runs in the cloud, the results go to its provider too.
• Tools can act. Some servers can write, delete or send. Review what tools a server exposes and whether the client asks before calling them.
• Trust the server. A server runs code on your machine. Install servers only from sources you trust.
• Prompt injection. Text returned by a tool can contain instructions aimed at the model. Good clients treat tool output as data.

In MediaFind

MediaFind includes an MCP server. In Settings → Connect to AI agents you can connect Claude, Claude Code or Cursor, and they can search your library, ask it questions and read transcripts on your computer. Assistants connected over MCP can also save to and search your Memory. Your files stay where they are; the assistant receives only the results it asks for, and if it is a cloud AI, those results go on to its provider. See the AI agents setup page and how the MCP server was built.

Frequently asked questions

Who created MCP?

MCP was introduced by Anthropic as an open standard in 2024 and is now supported by a range of AI apps and developer tools.

Does MCP send my files to the AI?

Not the files themselves. The assistant receives whatever the tool returns, such as search results or a transcript excerpt. If the assistant runs in the cloud, those results go to its provider.

What is the difference between an MCP server and an API?

An API is any programmatic interface. An MCP server is an API shaped to a shared standard, with tool descriptions a language model can read, so any MCP client can use it without custom code.