Installing an MCP Agent

An MCP agent is not magic.
It is a local server that Claude knows how to call.

If you understand that, 80% of bugs disappear.

1. Understand the Architecture

• The agent runs locally
• It exposes capabilities via MCP
• Claude Desktop acts as a client

Claude does not guess anything. Everything must be explicitly declared.

2. Prerequisites

• Claude Desktop installed
• Node.js or Python
• A dedicated folder per agent

Advice: 1 agent = 1 Git repo, versioned from the start.

3. Install and Launch the Agent

Node example:

npm install
npm run dev

You should see: "MCP server running..."
Otherwise: missing dependencies, port error, or misconfigured script.

4. Declare the Agent in Claude Desktop

In claude_desktop_config.json:

{
  "mcpServers": {
    "my-agent": {
      "command": "node",
      "args": ["index.js"]
    }
  }
}

Save then fully restart Claude Desktop.

5. Test and Debug

Common errors:

• Invalid JSON
• Wrong command
• Incorrect path
• Agent launched twice

Tip: systematically log the agent's inputs and outputs.

6. Security

An MCP agent can read, write, and execute code.

Strict rule: only run what you understand line by line.