GreyhavenHQ/greywall

What it solves

AI coding agents are powerful tools that often require broad access to your computer, which can create security risks. Greywall prevents these agents from accessing sensitive files like SSH keys or environment variables, and stops them from executing dangerous commands, all without the complexity of setting up containers.

How it works

Greywall acts as a security wrapper around your AI tools. It uses native operating system features (like Bubblewrap on Linux or Seatbelt on macOS) to create a restricted environment. By default, it denies access to your entire filesystem except for your current working directory and blocks all network traffic unless you explicitly allow it. It routes all network activity through a transparent proxy, providing a dashboard where you can see exactly what the agent is doing. You can use "Learning Mode" to automatically generate a custom profile based on the files an agent actually needs, or "Watch Mode" to observe an agent's behavior without any restrictions.

Who it’s for

It is designed for developers who use AI coding agents like Claude Code, Cursor, Aider, or Gemini and want to ensure these tools cannot accidentally or maliciously access private data, modify system files, or make unauthorized network requests.

Highlights

• Deny-by-default security: Blocks all filesystem and network access unless specifically permitted.
• Container-free: Uses native kernel-level sandboxing, avoiding the overhead of Docker.
• Built-in profiles: Includes pre-configured settings for popular AI agents and development toolchains.
• Learning mode: Automatically traces file access to generate least-privilege security profiles.
• Observability dashboard: Provides a live view of all network traffic and sandbox violations.
• Command blocking: Prevents destructive commands like rm -rf / from executing.