PrivacyGuard MCP

Private AI agent infrastructure for routing sensitive prompts to local LLMs.

What it does

• Classifies prompt text for sensitive keywords (PII/PHI/secrets)
• Routes sensitive prompts to local Ollama via MCP tool calls
• Emits JSON-RPC responses over stdio (MCP transport)
• Provides audit log retrieval from audit.log

Status

This project is an early MVP. It currently exposes MCP tools over stdio only.

Available tools:

• analyze_privacy
• route_to_local
• get_audit_logs

Prerequisites

# Rust toolchain
curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh

# Ollama (for local generation tool)
brew install ollama
ollama pull phi3:mini
ollama serve

Build

cargo build --release

Binary path:

./target/release/privacyguard-mcp

MCP client configuration example

{
  "mcpServers": {
    "privacyguard": {
      "command": "/absolute/path/to/target/release/privacyguard-mcp"
    }
  }
}

Local smoke test (stdio JSON-RPC)

./target/release/privacyguard-mcp <<'JSON'
{"jsonrpc":"2.0","id":1,"method":"initialize","params":{}}
{"jsonrpc":"2.0","id":2,"method":"tools/list","params":{}}
{"jsonrpc":"2.0","id":3,"method":"tools/call","params":{"name":"analyze_privacy","arguments":{"text":"My password is secret123"}}}
JSON

Development

cargo check
cargo test

Roadmap

• Replace keyword matching with stronger detection policies
• Add persistent consent workflow (grant/deny)
• Add integration tests for JSON-RPC/MCP behavior
• Add configurable policy and audit paths

License

MIT. See LICENSE.

Contributing

See CONTRIBUTING.md for development and PR guidelines.