Claude and Cursor MCP setup: a security-first guide

Updated 2026.

Adding MCP servers to Claude Desktop, Claude Code, or Cursor takes a few lines of JSON. Doing it safely takes a few more habits. Here is how to wire up servers without leaking secrets, over-exposing your filesystem, or quietly burning your context budget.

Where the config lives

Claude Desktop: claude_desktop_config.json
Cursor: .cursor/mcp.json
VS Code: .vscode/mcp.json or the mcp.servers setting
Project-level: a plain .mcp.json

Keep secrets out of the file

Do not paste an API token directly into the config. It leaks through the file and through git history. Use an environment variable and reference it.

{
  "mcpServers": {
    "github": {
      "type": "http",
      "url": "https://mcp.example.com",
      "headers": { "Authorization": "Bearer ${GITHUB_MCP_TOKEN}" }
    }
  }
}

Pin versions, do not auto-update

Running a server with npx -y package means you execute whatever was published most recently. Pin an exact version like package@1.4.2 and review updates before adopting them.

Scope filesystem access

A filesystem server pointed at /Users or $HOME lets the model read and write across your whole machine. Point it at the specific project directory instead.

Watch the server count

Every connected server loads its tool schemas into every request. If you have seven servers but use two, disable the other five. You will save context tokens and the model will pick tools more accurately.

Audit it in one command

Instead of checking all of this by hand, run a scanner over your config.

Free, local, zero-dependency

pipx install git+https://github.com/alih552/mcp-audit
mcp-audit

It auto-detects Claude, Cursor, VS Code, and Windsurf configs and flags every issue above, plus a token-cost estimate.

mcp-audit on GitHub