에이전트의 행동에,
검수할 수 있는
흔적을.

Runtrail is an independent macOS project by Woojin Lee. It is being developed to help a Mac user review the local activity of AI coding agents, initially Codex, while keeping their existing desktop workflow.

The proposed architecture uses an Endpoint Security system extension for process and file activity metadata, a Network Extension content filter for flow metadata, a protected local evidence broker, and a user-facing review app. The intended first release is observational: it does not automatically block activity. File contents, communication payloads, and authentication secrets are outside the default collection scope.

OS observations, agent context, attribution candidates, and human decisions are separate evidence categories. An observed file open does not prove that contents were read or exfiltrated; a network flow does not prove a successful connection or an HTTP request.

Development status: M0 prototype; no public build is available. Apple entitlement approval, appropriate distribution profiles, actual sensor deployment, permissions, and performance validation remain prerequisites. This page describes the product and design intent, not completed sensor validation.