Screen tool-call responses before the model trusts them.

import { Client } from "@modelcontextprotocol/sdk/client/index.js";

const SHIELD_API_KEY = process.env.SHIELD_API_KEY!;

async function safeCallTool(
  mcpClient: Client,
  toolName: string,
  args: Record<string, unknown>
): Promise<unknown> {
  // Call the MCP tool
  const result = await mcpClient.callTool({ name: toolName, arguments: args });
  const responseText = JSON.stringify(result);

  // Screen the response before the model sees it
  const res = await fetch("https://api.imposterhunter.com/v1/analyze", {
    method: "POST",
    headers: {
      "X-API-Key": SHIELD_API_KEY,
      "Content-Type": "application/json",
    },
    body: JSON.stringify({ input: responseText, context: "mcp_tool" }),
  });
  if (!res.ok) throw new Error(`Shield API ${res.status}`);
  const analysis = await res.json();

  if (analysis.action === "BLOCK") {
    // Log the detection, return safe error to the model
    console.warn("MCP Shield blocked:", {
      tool: toolName,
      category: analysis.primary_category_group,
      spans: analysis.evidence_spans,
    });
    throw new Error(`Tool response from ${toolName} was blocked by security policy.`);
  }

  return result; // Clean — pass to the model
}