MCP Server

Model Context Protocol server with comprehensive security features

Features

• PII/PCI Redaction: Automatic masking of credit cards, emails, phones, SSNs, addresses

• Field Allowlisting: Only safe fields returned to LLMs (blocks sensitive data)

• Free-Text Sanitization: Suspicious content detection and removal

• Size Limits: Payload size management and summarization (64KB max)

• Structured Responses: Consistent llm_view and meta format for audit trails

• Comprehensive Audit Logging: All requests logged with full context and client info

Security-First Design

1. Multi-Layer Defense: Authentication → Rate Limiting → PII Redaction → Field Filtering → Content Sanitization

2. Zero Trust Architecture: Every request is authenticated, rate-limited, and audited

3. Defense in Depth: Multiple security layers ensure no single point of failure

4. Fail-Safe Defaults: When in doubt, the server blocks access rather than allowing it

Real-World Security Examples

Before (Raw API Response):

After (LLM-Safe Response):

Project Structure

Installation

Prerequisites

• Python 3.11+

• AWS CLI configured (for Secrets Manager)

• Virtual environment

Setup

Configuration

Claude Desktop Integration

Add to your Claude Desktop MCP configuration:

Security Features

Authentication & Authorization

• AWS Secrets Manager Integration: API keys stored securely in AWS

• User Context: Each API key maps to a user with specific permissions

• Expiration Support: API keys can have expiration dates

• Default Key Fallback: Claude Desktop uses DEFAULT_API_KEY automatically

Rate Limiting

• Per-Endpoint Limits: Different limits for different tools

• Per-User Tracking: Rate limits applied per user ID

• Configurable Windows: Time windows (e.g., 1 hour, 1 day)

• In-Memory Storage: Fast rate limit checking

Audit Logging

• Comprehensive Logging: All requests logged with full context

• Security Events: Authentication failures, rate limit violations

• Structured Logs: JSON format for easy parsing

• File + Console: Logs to file and console for critical events

• Graceful Fallback: Console-only logging if file system is read-only

Data Protection & PII/PCI Redaction

The server automatically detects and masks sensitive data before it reaches the LLM, ensuring no PII/PCI data is exposed.

1. PII/PCI Masking Examples

What Gets Redacted/Masked:

• PAN/Credit Cards: 4111111111111111 → 4***-****-****-****1

• Bank Accounts: 1234567890 → 1*******0

• Emails: jane.doe@example.com → j***e@e***e.c***m

• Phones: 555-123-4567 → 5***-***-***7

• Addresses: 123 Main St → [address omitted]

• SSN/Tax IDs: 123-45-6789 → 1***-**-***9

• Personal Names: John Doe → J***n D***e

• Company Names: Acme Corporation → A***e C***********n

Raw vs. LLM-Safe Response:

2. Field Allowlisting Examples

Only approved fields are returned to prevent data leakage. Everything else is stripped.

Invoice Example:

3. Free-Text Sanitization Examples

Suspicious content in notes, descriptions, and comments is detected and sanitized.

4. Size Limits & Summarization Examples

Large payloads are automatically summarized to prevent token overruns.

5. Complete Security Response Example

Here's what a complete response looks like with all security features applied:

Security Configuration

The security profile is configured in security_profiles/finance-default.yaml:

Field Allowlists by Entity:

• Account: id, accountNumber, name, status, balance, currency, createdDate, updatedDate, type, industry

• Subscription: id, name, status, subscriptionStartDate, subscriptionEndDate, termType, autoRenew, renewalTerm, initialTerm, accountId, ratePlanId

• Invoice: id, invoiceNumber, status, amount, balance, dueDate, invoiceDate, currency, taxAmount, totalAmount, accountId, subscriptionId

Size Limits:

• Max Payload Size: 64KB

• Max Text Length: 200 characters

• Max List Items: 20 items

• Large lists are summarized: "980 more items omitted"

Meta Fields Explained

Response Meta Fields:

• requestId: Unique identifier for tracking requests

• userId: User who made the request (from API key)

• userRole: User's role (readonly, admin, etc.)

• apiKey: Masked API key for identification

• timestamp: When the request was processed

• fieldsRemoved: List of fields that were blocked/removed

• redactionProfile: Security profile used (e.g., "finance-default")

• securityApplied: Whether security enhancements were applied

• rateLimitInfo: Current rate limit status

• contentSanitized: Fields that had content sanitized

• listSummarized: Whether large lists were summarized

Security Event Logging:

Response Format

All responses follow this structure: