Real-time security filter for MCP tool communications. Works as a Claude Code hook to inspect all data sent to and received from MCP servers. Detects prompt injection, API key leakage, tool poisoning, ASCII smuggling, SSRF, Base64-encoded payloads, and Rug Pull attacks. Provides PASS/WARN/BLOCK decisions with per-server configuration, allowlisting, and logging. Zero dependencies, ~60ms latency.