How to integrate MCP plugins directly into VS Code

VS Code has built-in MCP client support — you don't need a separate extension to connect an MCP server to GitHub Copilot's agent mode. The mechanism is a single JSON config file, mcp.json, which you can place in one of two spots: .vscode/mcp.json inside your project (checked into source control so your whole team gets the same servers), or your user profile via the MCP: Open User Configuration command, which applies across every workspace you open.

The config file structure

Every server entry lives under a top-level servers object, keyed by a name you choose. A local server you spawn yourself uses "type": "stdio" with a command and args; a remote server you're only connecting to uses "type": "http" with a url. Both forms can coexist in the same file:

{
  "servers": {
    "github": {
      "type": "http",
      "url": "https://api.githubcopilot.com/mcp"
    },
    "playwright": {
      "command": "npx",
      "args": ["-y", "@microsoft/mcp-server-playwright"]
    }
  }
}

VS Code provides IntelliSense (autocomplete and validation) while you edit this file, since it ships a JSON schema for the format. The first time a configured server starts, VS Code shows a trust dialog — you have to explicitly confirm you trust the server and its capabilities before it runs, and individual tool calls it makes typically still need per-call confirmation in the Chat view unless you're running a sandboxed server on macOS/Linux, which gets auto-approval because it's already confined to a restricted environment.

Adding a server through the UI

Rather than hand-editing JSON, you can run the MCP: Add Server command from the Command Palette, which walks you through picking a server type and filling in the command or URL, then writes the entry into either the workspace or user mcp.json for you.

Once a server is trusted and running, its tools appear directly in Copilot Chat's agent mode — you don't invoke them by name, you just describe what you want and the model decides which registered tool to call, the same way it would call any built-in capability.