MCP Server Setup Skill

Set up MCP servers in the current workspace for testing with Claude Code.

Workflow

1. Identify the Source

Determine what the user has provided:

•Local path: A file path pointing to an existing MCP server on their system (e.g., /home/user/my-mcp-server, ./my-server)
•GitHub repo: A repository URL (e.g., https://github.com/user/mcp-server, github.com/user/mcp-server)
•npm package: A package name to run via npx (e.g., @modelcontextprotocol/server-filesystem)

2. Prepare the Server

For local paths:

•Verify the path exists and identify the server type (Node.js, Python, etc.)
•Check for package.json (Node.js) or Python entry point

For GitHub repos:

•Ask the user where to clone: "Where would you like me to clone this repository? (e.g., /home/claude/mcp-servers/ or a specific path)"
•Clone the repository to the specified location
•Install dependencies (npm install, pip install -r requirements.txt, etc.)

For npm packages:

•Use directly via npx in the configuration

3. Determine Server Configuration

Inspect the server to determine how to run it:

Server Type	Command	Args Example
Node.js (npm package)	`npx`	`["-y", "@package/name"]`
Node.js (local)	`node`	`["path/to/server.js"]`
TypeScript (local)	`npx`	`["tsx", "path/to/server.ts"]`
Python	`python` or `python3`	`["path/to/server.py"]`
Binary/Executable	Direct path	`["--flag", "value"]`

Check for:

•package.json → look for "main", "bin", or scripts
•*.py entry point → usually server.py, main.py, or __main__.py
•README for run instructions
•Environment variables required (check .env.example, README, or code)

4. Create .mcp.json

Create or update .mcp.json in the project root (current working directory):

json

{
  "mcpServers": {
    "server-name": {
      "command": "node",
      "args": ["/absolute/path/to/server.js"],
      "env": {
        "API_KEY": "${API_KEY}",
        "DEBUG": "true"
      }
    }
  }
}

Configuration rules:

•Use absolute paths for reliability
•command: The executable to run
•args: Array of command-line arguments
•env: Environment variables (supports ${VAR} and ${VAR:-default} expansion)
•Server name should be descriptive and lowercase (e.g., my-api-server, github-tools)

If .mcp.json already exists:

•Read existing content
•Merge new server into mcpServers object
•Preserve existing server configurations

5. Remind User to Restart

After setup, always inform the user:

Important: Claude Code must be restarted for the MCP server to load. Please restart your Claude Code session (close and reopen, or use the restart command) to activate the new MCP server. After restarting, you can verify the server is loaded by using /mcp command.

Common Patterns

Node.js server from local path

json

{
  "mcpServers": {
    "my-server": {
      "command": "node",
      "args": ["/home/user/my-mcp-server/dist/index.js"],
      "env": {}
    }
  }
}

TypeScript server (uncompiled)

json

{
  "mcpServers": {
    "my-ts-server": {
      "command": "npx",
      "args": ["tsx", "/home/user/my-mcp-server/src/index.ts"],
      "env": {}
    }
  }
}

Python server

json

{
  "mcpServers": {
    "python-server": {
      "command": "python3",
      "args": ["/home/user/mcp-server/server.py"],
      "env": {
        "PYTHONPATH": "/home/user/mcp-server"
      }
    }
  }
}

npm package via npx

json

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/allowed/path"],
      "env": {}
    }
  }
}

Troubleshooting Guidance

If the user reports issues after restart:

•"Connection closed" errors: Check command path, ensure dependencies installed
•Server not appearing: Verify .mcp.json is in project root and syntax is valid
•Permission errors: Check file permissions on server executable
•Missing dependencies: Run npm install or pip install in server directory

Suggest running /mcp after restart to check server status.