MCP Client

An MCP client is the side that calls up an MCP server to borrow the functions inside it. If MCP (Model Context Protocol) is the phone line, the server is the one on the other end actually drawing pictures or digging through files, and the client is the one that talks with the user and places the call to get that work done.

For example, if a program like Claude Desktop, Cursor, or Codex connects to the graphics tool ComfyUI, then Claude Desktop or Cursor is playing the role of the MCP client. When a user types "run this video model locally" into the chat window, the client passes that request to the server, and the server actually checks the hardware and downloads the files.

In other words, an MCP client isn't the side that builds new functionality itself — it's more like a doorway that pulls an existing server's functions into the chat window. Even with the same server, connecting to a different client changes the screen the user sees and how the conversation feels.

The article writes that agents like Claude Desktop, Claude Code, Cursor, and Codex "connect directly to ComfyUI already installed on my computer" to build workflows. Here, programs that talk with the user and send requests to the server — like Claude Desktop or Cursor — are the MCP clients. It's easy to confuse this with ComfyUI itself (the side providing the tools), but the distinction is simple: the side lending out tools is the server, and the side borrowing them is the client.

Open a program that supports MCP (e.g., Claude Desktop) and add an MCP server address or local connection info in the settings. Then try typing this into the chat:

"Set up a local ComfyUI MCP connection, here's the install guide: [doc URL]"

Once connected, you can check whether the client is properly pulling information from the server by asking something like "show me the list of installed nodes."