Connecting Tools

How MCP Connection Works

When you connect an MCP tool, your AI client starts a local process (usually via npx) that runs the MCP server. The client communicates with this server over stdio using the JSON-RPC protocol. No data leaves your machine unless the tool explicitly makes network requests.

Each AI client stores its MCP configuration in a JSON file. You add server entries to the mcpServers object, and the client manages starting and stopping them automatically.

Claude Desktop

Claude Desktop reads MCP server configuration from a JSON file on your machine. The file location depends on your operating system:

Step-by-step setup

1. 1Open Claude Desktop and go to Settings → Developer.
2. 2Click “Edit Config” to open the config file in your editor. If the file doesn't exist, create it at the path above.
3. 3Add the MCP server configuration. Here's an example using the Filesystem server:

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

4. 4Save the file and fully restart Claude Desktop (quit and reopen, not just close the window).
5. 5Look for the hammer icon in the chat input — clicking it should show your connected tools.

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/you/Documents"]
    },
    "github": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": {
        "GITHUB_TOKEN": "ghp_your_token_here"
      }
    }
  }
}

Cursor

Cursor supports MCP tools through a per-project configuration file in your project root.

1. 1Create a .cursor directory in your project root if it doesn't exist.
2. 2Create mcp.json inside it with your server configuration:

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

3. 3Restart Cursor. The tools will be available in your Composer and Agent panels.

Windsurf

Windsurf (by Codeium) uses a global MCP configuration file.

1. 1Open Windsurf and go to Settings → MCP (or create the file manually at the path above).
2. 2Add your MCP server configuration:

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

3. 3Restart Windsurf. The tools will appear in Cascade's tool list.

OpenClaw

OpenClaw is an AI agent orchestration platform with 40+ communication channels and its own skills/plugins system. It uses a global MCP configuration file.

1. 1Create the ~/.openclaw directory if it doesn't exist.
2. 2Open or create openclaw.json and add your MCP server configuration:

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

3. 3Restart OpenClaw. Your MCP tools will be available alongside OpenClaw's native skills and plugins.

Other Clients

Any MCP-compatible client can connect to tools from Hive Market. The core configuration is the same JSON structure — you just need to know where your client reads it from.

The minimal configuration for any MCP server looks like this:

{
  "mcpServers": {
    "server-name": {
      "command": "npx",
      "args": ["-y", "@scope/package-name"],
      "env": {}
    }
  }
}

• command — the executable to run (usually npx, node, or uvx for Python servers)
• args — arguments passed to the command (-y auto-confirms npx prompts)
• env — environment variables for the server process (API keys, tokens, etc.)

Troubleshooting