Show HN: A police department for your Claude Code agents (github.com)

A new tool, agent-pd, has been introduced to enhance the security and auditing of Claude Code agents by providing a comprehensive logging system that records all interactions and permissions related to tool usage. Rather than interfering with agent operations, agent-pd functions as a passive observer—capturing every event without blocking operations. This logging includes a hash-chained audit log and the capability to replay sessions, ensuring that any violations, such as unauthorized actions or out-of-scope accesses, are documented and easily accessible for review. This tool is significant for the AI/ML community as it raises operational accountability, allowing developers to monitor agent behavior in real-time and retrospectively analyze any deviations from expected conduct. The system employs six deterministic detectors to identify critical permission bypasses and other potential security threats while generating minimal overhead by requiring zero token costs. By improving visibility into agent actions and maintaining an accurate record of activities, agent-pd not only enhances security but also builds user trust in leveraging AI systems for complex tasks.