98% of What Claude Code Does, I Never See. The 0.27% It Stops for Is the Point (grith.ai)

A new open-source project called grith has been introduced as a security proxy for AI coding agents, operating at the operating system level. Grith intercepts system calls made by coding tools like Claude Code and Codex, assessing actions such as file reads, process spawning, and network connections before they are executed. This monitoring drastically reduces the burden on developers to approve each action—allowing about 98.4% of operations to pass without intervention while flagging only 0.27% for human review. This is crucial because it protects sensitive resources (like SSH keys and AWS logins) from unauthorized access by the AI agent, which could inadvertently adopt unsafe behaviors. The significance of grith lies in its ability to enhance security without hindering workflow. Unlike traditional permission prompts that can overwhelm developers, grith operates quietly in the background, only alerting users to potentially risky actions. It employs deterministic matching for enforcement, ensuring that malicious inputs designed to manipulate the AI cannot evade its oversight. By combining high efficiency with robust security, grith represents a meaningful advancement in the safe deployment of AI coding agents, enabling unattended operations without jeopardizing sensitive data.