How to Set Up Claude Desktop and Connect MCP Servers

What Claude Desktop is and why MCP matters

Claude Desktop is Anthropic's native desktop application for macOS and Windows. It is a chat client that runs as its own program instead of a browser tab, which is what lets it talk to tools on your machine.

That tool access comes from MCP, the Model Context Protocol. MCP is an open standard for connecting an assistant to external systems: your filesystem, a database, a git repo, an API, or a retrieval index. Each connection is an MCP server, a small program that exposes a set of tools the model can call. Without MCP, Claude Desktop is a good chat window. With MCP, it can read your files, query your data, and act on real systems while you stay in one conversation.

This guide covers the setup that actually gets servers running: installing the app, finding and editing the config file, adding a server, and verifying the connection. It does not cover plans or pricing, which change and vary by account. Check your Claude account for what your plan includes.

Install Claude Desktop and open the config

Download the app for your operating system, install it, and sign in with your Claude account. The desktop app is separate from the web version and separate from Claude Code, the terminal tool. You can run all three; they share your account, not their configuration.

Add your first MCP server

Every server is an entry inside an mcpServers object. Each entry has a name you choose, a command to run, and an args array passed to that command. Here is a minimal config that connects the filesystem server and scopes it to one folder:

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-filesystem",
        "/Users/you/projects"
      ]
    }
  }
}

To run more than one server, add more keys to the same object. Keep the JSON valid: commas between entries, no trailing comma after the last one, and double quotes everywhere. A single syntax error stops the whole file from loading, so paste it through a JSON linter if servers stop appearing.

Some servers need secrets, like an API token. Those go in an env object on the server entry, not in args, so keys stay out of your command line:

{
  "mcpServers": {
    "example": {
      "command": "npx",
      "args": ["-y", "some-mcp-server"],
      "env": { "API_TOKEN": "your-token" }
    }
  }
}

Save the file. The app does not hot-reload it.

Verify the connection and troubleshoot

Claude Desktop reads the config only at startup. Quitting the window is not enough on either OS, because the app keeps running in the tray or menu bar. Quit fully, then reopen.

After restart, open a chat and look for the tools indicator, usually a slider or tool icon near the input box. Click it to see the connected servers and the tools each exposes. Ask the model to use one, for example "list the files in my projects folder," and approve the tool call when prompted. If the tools show up and the call runs, the server is wired correctly.

When a server does not appear, work through this order. First, validate the JSON. Second, confirm the runtime exists by running the exact command from a terminal yourself; npx -y @modelcontextprotocol/server-filesystem /path should start without an error. Third, use an absolute path for the command if the app cannot find it on PATH, which happens on macOS when the app launches with a minimal environment. Fourth, read the MCP logs, which Claude Desktop writes to a logs folder next to the config file. The log names the server that failed and usually the reason.

Connect an RDK retrieval MCP to cut token spend

Once the app talks to tools, the highest-value server to add is a retrieval index. Here is the problem it solves. A desktop agent answers from the model every time, which means you pay to regenerate the same facts about your own notes, docs, and decisions on every session. That is slow and it is wasteful.

RDK is a retrieval layer built for this. You index your local vault, whether that is Obsidian notes, project docs, or code, as encrypted private chunks. You expose that index to Claude Desktop as an MCP server, declared in the same mcpServers object as any other. Once connected, the desktop agent searches your index before it calls the model. A settled answer that already exists in your vault comes back from retrieval instead of being regenerated token by token.

The savings come from where the work lands. A private vault answers 40 to 65 percent of everyday queries directly, the public RDK network adds another 15 to 20 percent, and the model handles only the remaining 5 to 10 percent as fallback. For a desktop agent you use all day against the same knowledge base, that is a large cut in token spend and a faster response, because retrieval is cheaper than inference. Add it as one more server entry, restart, and the desktop agent starts answering from what you already know.