How to Install an MCP Server in Claude, Cursor, Windsurf, and VS Code

TL;DR: Every MCP client installs servers the same way at heart — you add a small JSON block naming a command to run, then restart or refresh the client. This guide covers the universal config shape and then the exact file location and steps for Claude Desktop, Claude Code, Cursor, Windsurf, and VS Code, plus how to confirm your tools actually loaded.

The Model Context Protocol (MCP) lets AI clients launch small helper programs — "servers" — that expose tools, resources, and prompts the model can use. Once you have installed one server, you have basically installed them all: the mechanics barely change from client to client. The friction is almost always in two places — where the config file lives, and remembering to restart — so this guide focuses on getting both right.

What you need before you start

Most MCP servers are distributed as npm or Python packages and launched on demand, so you need one of two runtimes on your machine:

• Node.js (which provides npx) for JavaScript/TypeScript servers.
• Python with uv (which provides uvx) for many Python servers.

You do not usually install the server package yourself — npx -y or uvx fetches and runs it the first time your client launches it. You just need the runtime available on the same PATH your client uses, which is the single most common source of setup problems on macOS and Windows.

The universal config shape

Nearly every client describes a server with three fields: the command to run, its args, and an optional env object for secrets. Here is the canonical stdio block:

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/path/to/allowed/dir"],
      "env": {}
    }
  }
}

Keep secrets in env, never inline in args, and give each server a short, unique name (the key). Once you understand this block, the only per-client differences are the file it goes in and how you reload it.

Claude Desktop

Claude Desktop reads a JSON file whose fastest access is Settings → Developer → Edit Config, which creates the file if it does not exist. The paths are:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

Add your server under mcpServers, save, then fully quit and reopen Claude Desktop — closing the window is not enough, because servers launch only at startup. If a server shows zero tools, read the per-server log at ~/Library/Logs/Claude/mcp*.log (macOS). The full walkthrough lives on the Claude Desktop setup guide.

Claude Code

Claude Code is CLI-first, so you rarely touch a config file by hand. Add a server with:

claude mcp add filesystem -- npx -y @modelcontextprotocol/server-filesystem /path/to/allowed/dir

Everything after -- is the launch command. Use -e KEY=value for secrets, and --scope project to write a shared .mcp.json your team commits (or --scope user to make it global). Confirm it worked with claude mcp list, or the /mcp command inside a session. See the Claude Code setup guide for scope details.

Cursor

Cursor reads two files with the same mcpServers shape: ~/.cursor/mcp.json for every project, and .cursor/mcp.json in a project root to scope servers to that repository. Add servers from Cursor Settings → MCP → Add new MCP server, or edit the JSON directly. Make sure the server is toggled on, and remember that MCP tools are called from the Agent (Composer) — not the plain chat. Committing .cursor/mcp.json is the easiest way to share servers with a team. Full steps are on the Cursor setup guide.

Windsurf

Windsurf (Codeium) keeps its config at ~/.codeium/windsurf/mcp_config.json, again using the mcpServers shape. Edit it from the MCP settings inside Cascade or open the file directly, then use the refresh action in the MCP panel so Windsurf picks up the change without a full restart. The Windsurf setup guide has the details.

VS Code (GitHub Copilot)

VS Code is the one client that breaks the pattern. It nests servers under a servers key — not mcpServers — in .vscode/mcp.json, and supports an inputs array so secrets are prompted for instead of hard-coded:

{
  "servers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/path/to/allowed/dir"]
    }
  }
}

Run MCP: Add Server from the Command Palette to scaffold it, then use the tools from Copilot Chat in Agent mode. If you copy a config from any other client, rename mcpServers to servers. The VS Code setup guide covers the rest.

Verifying it worked (and fixing it when it did not)

After adding a server, the checklist is the same everywhere:

1. Did the client pick it up? Look for the server in the client's MCP panel or tool list.
2. Does it show tools, or zero tools? Zero tools means the process started but errored before responding — check for a missing env var, a command not on PATH, or stray stdout output.
3. Did you restart or refresh? GUI clients need a full restart; panel-based clients need a refresh.

If it is still broken, the client's per-server MCP log almost always contains the real error text, which maps directly to one of those causes. For a deeper walkthrough of every failure mode, see our post on why an MCP server won't connect.

Where to find servers to install

Once you know how to install one, the fun part is deciding which to add. Browse the AllMCPs directory by category, or start from a ranked, use-case guide — the best MCP servers for databases, for GitHub and version control, or for web search. Every listing shows the exact command and arguments to paste into the config for your client. And if you want the setup steps for a specific client in one place, the MCP client setup guides have you covered.