Expose internal MCPs to Claude without leaking PII
Wann einsetzen: Your internal tools (postgres, CRM) would leak customer names / emails / SSNs into LLM context if used directly.
Voraussetzungen
pip install mcp-gateway[presidio]— presidio extra pulls spaCy models
Ablauf
-
Wire upstream MCPs through gatewayConfigure mcp-gateway with upstream postgres MCP; enable the Presidio plugin with redaction for EMAIL, PHONE, US_SSN.✓ Kopiert→ Config file with plugin chain
-
Test redactionRun a query returning customer emails; confirm the gateway replaces them with <EMAIL> placeholders before reaching the LLM.✓ Kopiert→ Claude sees only placeholders
Ergebnis: LLM-usable tool access without raw PII crossing the boundary.
Fallstricke
- Over-redaction breaks downstream joins — Use consistent pseudonyms (Presidio supports hash-based) so the LLM can still reason across redacted fields