Behavioral Audit MCP

Check a draft support reply for what it should say and doesn't, before it goes to the customer.

Part of Agentic Diaries. One hosted tool, disclosure_check.

What it does

Give it what a customer said and the reply you are about to send. It flags any should-know policy the customer's situation made relevant that your draft failed to proactively surface: a refund window, an auto-renewal, a retention offer, a warranty term. The knew-but-did-not-say gap, on a single message.

It returns a verdict, pass or gap. On a gap, it names each missed policy and the line that should have been there. It detects and suggests. It does not rewrite your reply, send anything, or keep the message beyond the check.

Related MCP server: @aws/ftr-eval-mcp

Install

One line. No login, no signup.

Claude Code:

claude mcp add --transport http behavioral-audit https://receipts-mcp.vercel.app/api/mcp

If you add it mid-session, restart Claude Code (or run /mcp) so the new tool appears.

Claude Desktop (or any MCP client), add to your config:

{
  "mcpServers": {
    "behavioral-audit": {
      "type": "http",
      "url": "https://receipts-mcp.vercel.app/api/mcp"
    }
  }
}

The tool: disclosure_check

Input:

• customer_message (string) , what the customer said, verbatim.

• draft_reply (string) , the reply you are about to send.

Output: a verdict plus, on a gap, the policies your draft left out.

Example

A customer asks to cancel. Your draft confirms it but never mentions the refund window they still qualify for.

Request:

{
  "customer_message": "I want to cancel my subscription today.",
  "draft_reply": "Done, your subscription is cancelled. Sorry to see you go."
}

Result:

{
  "verdict": "gap",
  "misses": [
    {
      "id": "refund_window",
      "title": "Refund window",
      "suggestion": "You are within your 14-day refund window, so you can still request a refund for this billing period.",
      "reason": "The customer is cancelling and qualifies for a refund they were not told about."
    }
  ]
}

A clean reply returns verdict: "pass" with an empty misses list. That means clear to send, not that nothing ran.