MCP Conceal

An MCP proxy that pseudo-anonymizes PII before data reaches external AI providers like Claude, ChatGPT, or Gemini.

sequenceDiagram
    participant C as AI Client (Claude)
    participant P as MCP Conceal
    participant S as Your MCP Server
    
    C->>P: Request
    P->>S: Request
    S->>P: Response with PII
    P->>P: PII Detection
    P->>P: Pseudo-Anonymization
    P->>P: Consistent Mapping
    P->>C: Sanitized Response

MCP Conceal performs pseudo-anonymization rather than redaction to preserve semantic meaning and data relationships required for AI analysis. Example: john.smith@acme.com becomes mike.wilson@techcorp.com, maintaining structure while protecting sensitive information.

Installation

Download Pre-built Binary

1. Visit the Releases page
2. Download the binary for your platform:

3. Make executable: chmod +x mcp-server-conceal-* (Linux/macOS)
4. Add to PATH:
   • Linux/macOS: mv mcp-server-conceal-* /usr/local/bin/mcp-server-conceal
   • Windows: Move to a directory in your PATH or add current directory to PATH

Building from Source

git clone https://github.com/gbrigandi/mcp-server-conceal
cd mcp-server-conceal
cargo build --release

Binary location: target/release/mcp-server-conceal

Quick Start

Prerequisites

Install Ollama for LLM-based PII detection:

1. Install Ollama: ollama.ai
2. Pull model: ollama pull llama3.2:3b
3. Verify: curl http://localhost:11434/api/version

Basic Usage

Create a minimal mcp-server-conceal.toml:

[detection]
mode = "regex_llm"

[llm]
model = "llama3.2:3b"
endpoint = "http://localhost:11434"

See the Configuration section for all available options.

Run as proxy:

mcp-server-conceal \
  --target-command python3 \
  --target-args "my-mcp-server.py" \
  --config mcp-server-conceal.toml

Configuration

Complete configuration reference:

[detection]
mode = "regex_llm"                # Detection strategy: regex, llm, regex_llm
enabled = true                    
confidence_threshold = 0.8        # Detection confidence threshold (0.0-1.0)

[detection.patterns]
email = "\\b[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+\\.[A-Z|a-z]{2,}\\b"
phone = "\\b(?:\\+?1[-\\.\\s]?)?(?:\\(?[0-9]{3}\\)?[-\\.\\s]?)?[0-9]{3}[-\\.\\s]?[0-9]{4}\\b"
ssn = "\\b\\d{3}-\\d{2}-\\d{4}\\b"
credit_card = "\\b\\d{4}[-\\s]?\\d{4}[-\\s]?\\d{4}[-\\s]?\\d{4}\\b"
ip_address = "\\b(?:(?:25[0-5]|2[0-4][0-9]|[01]?[0-9][0-9]?)\\.){3}(?:25[0-5]|2[0-4][0-9]|[01]?[0-9][0-9]?)\\b"
url = "https?://[^\\s/$.?#].[^\\s]*"

[faker]
locale = "en_US"                  # Locale for generating realistic fake PII data
seed = 12345                      # Seed ensures consistent anonymization across restarts
consistency = true                # Same real PII always maps to same fake data

[mapping]
database_path = "mappings.db"     # SQLite database storing real-to-fake mappings
retention_days = 90               # Delete old mappings after N days

[llm]
model = "llama3.2:3b"             # Ollama model for PII detection
endpoint = "http://localhost:11434"
timeout_seconds = 180
prompt_template = "default"       # Template for PII detection prompts

[llm_cache]
enabled = true                    # Cache LLM detection results for performance
database_path = "llm_cache.db"
max_text_length = 2000

Configuration Guidance

Detection Settings:

• confidence_threshold: Lower values (0.6) catch more PII but increase false positives. Higher values (0.9) are more precise but may miss some PII.
• mode: Choose based on your latency vs accuracy requirements (see Detection Mode

... View full README on GitHub