Guard sits in front of your MCP server and runs 22 deterministic checks across five layers on every call. It blocks tool poisoning, rug pulls (CVE-2025-54136), and hidden Unicode payloads in tool descriptions before they land. No AI model does the checking — same verdict every time, no token bill. Point your agent at Guard instead of building this yourself. Three ways to run the same engine: hosted gateway, embedded SDK, or sidecar CLI for Cursor / Claude Code / Copilot stdio servers.