Complete the Claude Desktop MCP server configuration fields

Claude Desktop discovers MCP servers from a single file named `claude_desktop_config.json`. You edit it through Settings, and it lives in the app's support directory on your machine. Inside that file, every server you want to use goes under one top-level key: `mcpServers`. Each entry is named by you, and its value is an object describing how to reach that server. For a **stdio server** (a local program), you give two fields. The `command` field names the executable to run, like `npx` or `python`. The `args` field is an array of arguments passed to that command, such as the package name or a script path. For a remote server, you do not launch anything locally. Instead you give a `url` field pointing at the server's Streamable HTTP endpoint. Claude Desktop connects over the network rather than spawning a process. That single distinction, run a command versus hit a URL, is the whole answer.

Claude Desktop reads one JSON file at startup: `claude_desktop_config.json`. It sits in the per-user application support directory, and you usually edit it through the Settings dialog. The file has one top-level key that matters here: `mcpServers`. It maps a server name you choose to a connection spec. That spec branches on transport. For a stdio server, Claude Desktop launches a local subprocess and talks JSON-RPC over its standard input and output. You specify `command`, the executable such as `npx`, and `args`, an array passed to it such as the server package. An optional `env` object injects environment variables, which is where secrets like API keys belong. For a remote server using Streamable HTTP, there is no local process to spawn. You supply a `url` field pointing at the server's HTTP endpoint, and Claude Desktop connects over the network instead. So the shape of an entry tells you the transport. The pair command plus args means stdio, while url means remote HTTP. Get those field names right and the host wires the discovered tools into the model automatically.