Cupcake is a free AI security tool designed to enforce policies for AI agents, ensuring they adhere to predefined rules while enhancing performance. It operates as a policy enforcement layer that intercepts agent actions, evaluates them against user-defined rules, and provides feedback or corrections as necessary. By utilizing Open Policy Agent (OPA) and Rego for rule definitions, Cupcake allows for deterministic rule-following without consuming model context, making it efficient and effective for various coding environments.

The application supports multiple AI coding agents, including Claude Code, Cursor, Factory AI, and OpenCode, with native integrations and clear policy separation for each harness. It offers features such as granular tool control, observability through structured logs, and a unique LLM-as-Judge capability, allowing for dynamic evaluation of actions. Cupcake aims to provide robust governance for AI agents, making it a valuable tool for organizations focused on AI safety and compliance.