mcp-server-starter

A small but real MCP server (FastMCP) — with a destructive action guarded behind human confirmation.

Most "MCP server" examples stop at a hello_world tool. This one shows the part that actually matters when an agent is on the other end: a destructive tool that refuses to run until a confirmation token is supplied, forcing the agent to surface the decision to a human instead of wiping data on its own.

It exposes a tiny inventory domain as MCP tools + one resource, over stdio transport — drop-in for Claude Desktop, Cursor, or the MCP Inspector.

Tools

The guard pattern (delete_all_items) is the point: call it once with no token → you get back {"status": "confirmation_required", "confirm_token": "..."} → a human approves → call again with the token → it executes. An irreversible action never runs on the agent's say-so alone.

Related MCP server: mcp-server-demo

Quick start

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

python server.py            # stdio transport

Inspect it interactively:

npx @modelcontextprotocol/inspector python server.py

Wire it into Claude Desktop / Cursor

Add to your MCP config (claude_desktop_config.json or .cursor/mcp.json):

{
  "mcpServers": {
    "inventory-starter": {
      "command": "python",
      "args": ["/absolute/path/to/mcp-server-starter/server.py"]
    }
  }
}

Tip (Claude Code schema): for HTTP-based servers use "type": "http" — the validator rejects the spec-name "streamable-http" even though they're the same transport. (stdio servers like this one don't need a type.)

Going further

• Swap the in-memory _ITEMS dict for Postgres/Supabase (see the companion fastapi-supabase-multitenant-starter).

• Route the confirmation to Slack/web instead of a token — see agent-guardrails for a pluggable approver + audit trail.

• For a LangGraph agent that consumes a guarded tool like this, see langgraph-hitl-agent.

License

MIT.