Remote URL vs a local MCP command

Every MCP listing is one of two shapes. Remote means the publisher already hosts a URL. Local means the agent starts a process on your machine. The install row on a listing shows that value first, then the client buttons.

Remote URL

Looks like https://…/mcp or an SSE path. Cursor, Claude Code, VS Code, Windsurf, ChatGPT, and Codex can all point at it. Nothing runs on your laptop except the agent. Auth, if any, is a header the publisher documents.

1{
2  "mcpServers": {
3    "github": {
4      "url": "https://api.githubcopilot.com/mcp/"
5    }
6  }
7}

PlugMCP probes remote URLs every fifteen minutes and shows uptime on the listing. A 401 is not downtime. It means the server is up and wants you to sign in.

Local command

Looks like npx, uvx, or docker run. The registry lists a package. Your agent starts that process over stdio. Filesystem, browser, and many GitHub community servers work this way. There is no public URL to copy.

1{
2  "mcpServers": {
3    "memory": {
4      "command": "npx",
5      "args": ["-y", "@modelcontextprotocol/server-memory"]
6    }
7  }
8}

• npm → npx -y <package>
• PyPI → uvx <package>
• OCI → docker run -i --rm <image>

Required environment variables appear under the command on the listing. Replace <secret> with your own key. Do not paste those values into a chat.