@letsping/mcp

MCP server that lets Claude Desktop, Cursor, and other MCP clients route risky tool calls through LetsPing for human approval.

Usage

Zero-Config (via npx)

You can run the server directly without installing it, as long as you have your API Key.

Claude Desktop Config (claude_desktop_config.json):

{
  "mcpServers": {
    "letsping": {
      "command": "npx",
      "args": ["-y", "@letsping/mcp"],
      "env": {
        "LETSPING_API_KEY": "lp_sk_..."
      }
    }
  }
}

What this looks like in practice

After you add the server, your agent gets a tool named ask_human.

• Call it with { service, action, payload } and a request appears in the LetsPing dashboard.
• The human can approve, reject, or patch fields.
• The tool returns JSON text with a status field and the executed_payload that should run.

Tools Provided

ask_human

Request approval or input from a human operator.

Arguments:

• service (string): The name of your agent (e.g. "billing-bot").
• action (string): What you are trying to do (e.g. "refund-user").
• payload (object): The data needing review (e.g. { "amount": 50, "user_id": "123" }).
• priority (string, optional): "low" | "medium" | "high" | "critical". Defaults to "medium".
• role (string, optional): Who should approve this? (e.g. "finance", "devops").
• timeout (number, optional): Max wait time in milliseconds (default ~24h).

Return shape (content text JSON):

// APPROVED, no modifications
{
  "status": "APPROVED",
  "executed_payload": { /* final payload */ }
}

// APPROVED, with human modifications
{
  "status": "APPROVED_WITH_MODIFICATIONS",
  "message": "The human reviewer authorized this action but modified your original payload. Please review the diff_summary to learn from this correction.",
  "diff_summary": { "changes": { /* field-level diff */ } },
  "original_payload": { /* as requested by the agent */ },
  "executed_payload": { /* what will actually run */ }
}

// REJECTED
{
  "status": "REJECTED",
  "message": "The human operator rejected this action. Do not proceed with the plan.",
  "metadata": { /* optional audit fields */ }
}

How an MCP agent should behave:

• Never proceed with the risky action if status === "REJECTED".
• If status === "APPROVED_WITH_MODIFICATIONS", prefer executed_payload over the original; optionally learn from diff_summary.
• If status === "APPROVED", proceed using executed_payload.

When you should not use this

• You already have direct SDK integration and do not use MCP. In that case prefer @letsping/sdk or the framework adapters.
• You need arbitrary tools exposed over MCP. This server only exposes ask_human for approvals, not a general tool registry.
• You want to stream full token logs. The focus here is governing tool calls and payloads, not transcript capture.

Development

1. Clone the repo.
2. Run pnpm install.
3. Run pnpm build.