Protecting MCP Servers

Model Context Protocol servers often connect assistants to sensitive systems such as CRMs, document stores, and ticketing tools. FastPII helps you keep those tool boundaries clean.

Basic MCP server setup

from fastpii import PrivacyGuard
from fastpii.integrations.mcp import MCPServer

guard = PrivacyGuard(regions=["cz"])
server = MCPServer(regions=["cz"])

Expose a sanitization tool

from fastpii import PrivacyGuard
from fastpii.integrations.mcp import MCPServer

guard = PrivacyGuard(regions=["cz"])
server = MCPServer(regions=["cz"])


def sanitize_text(text: str) -> str:
    return guard.redact(text)


def detect_text(text: str):
    result = guard.detect(text)
    return result.findings

Use one tool for transformation and another for inspection so clients can choose whether they need findings, sanitized text, or both.

Tool boundary pattern

Apply FastPII in two places:

1. Before the server forwards user content to a backend system.
2. Before the server returns backend content to the MCP client.

from fastpii import PrivacyGuard

guard = PrivacyGuard(regions=["cz"])


def sanitize_request_payload(payload: str) -> str:
    return guard.remove(payload)


def sanitize_response_payload(payload: str) -> str:
    return guard.redact(payload)

Claude Desktop integration pattern

When connecting the MCP server to Claude Desktop, register it as a privacy-focused utility server and route sensitive document handling through the FastPII-backed tools first.

Recommended flow:

• Claude Desktop sends a document snippet to your MCP server.
• The server runs detect() or redact().
• Only sanitized text is forwarded to downstream tools or returned to the assistant.

Good defaults

• Use redact() for server responses.
• Use remove() for backend actions that do not need PII.
• Keep the raw source system separate from assistant-facing tool output.