Many people are connecting the Model Context Protocol (MCP) while using Cursor or Claude Code these days, right? It's fun to directly connect various tools to your AI assistant, but you often get stuck with unexplained bugs during the initial setup. Here are three tips to help beginners reduce the struggle when setting things up themselves.

First, when connecting a remote MCP server in Cursor, if you specify mcp.json in your "type": "streamable-http" file, there's a bug where the CLI agent fails to recognize this setting and omits the file entirely. In these cases, try changing the type simply to http like this haha.

{
  "mcpServers": {
    "my-remote-server": {
      "type": "http",
      "url": "https://example.com/mcp"
    }
  }
}

Also, when some third-party adapters fail to register because they are set to "type": "local", you can manually change them to "type": "stdio" and it will work just fine.

Second, Claude Code is much stricter about syntax and format validation. If there's a typo in your configuration or the schema doesn't match, it will throw errors immediately via terminal diagnostic commands like \doctor or \mcp, so it's a good idea to check those. On the bright side, Claude Code supports environment variable substitution in the ${VAR} format to prevent exposing secret keys, which is useful when sharing configurations in collaborative projects. Note that using Streamable HTTP is now officially recommended over SSE.

Lastly, if you're using Cursor, don't keep too many MCP servers enabled at once. There's a limitation where exceeding 40 tools can cause display issues in the UI or lead to dropped features. Since you also have to manually enable them in the chat window every time you restart the editor, it's best to just keep the essential servers enabled to stay efficient.