Greywall — Sandbox for AI Coding Agents

Greywall is a container-free, deny-by-default sandbox for AI agents on Linux and macOS. It restricts filesystem access, network connections, and system calls to only what you explicitly allow, so tools like Claude Code, Cursor, Codex, and other AI coding agents can't access your SSH keys, environment secrets, or anything outside the working directory.

Use --learning to trace what a command needs and auto-generate a least-privilege config profile. All network traffic is transparently redirected through greyproxy, a deny-by-default transparent proxy with a live allow/deny dashboard.

Supports Linux and macOS. See platform support for details.

• Deny-by-default filesystem — only the working directory is accessible unless you allow more
• Network isolation — all traffic blocked or routed through greyproxy with a live dashboard
• Command blocking — dangerous commands like rm -rf / and git push --force are denied
• Built-in agent profiles — one-command setup for Claude Code, Cursor, Codex, Aider, Goose, Gemini, OpenCode, Amp, Cline, Copilot, and more
• Learning mode — traces filesystem access and auto-generates least-privilege profiles
• Five security layers on Linux — Bubblewrap namespaces, Landlock, Seccomp BPF, eBPF monitoring, TUN-based network capture
• No containers required — kernel-enforced sandboxing without Docker overhead

# Sandbox a command (network + filesystem denied by default)
greywall -- curl https://example.com

# Sandbox an AI coding agent with a built-in profile
greywall -- claude

# Learn what filesystem access a command needs, then auto-generate a profile
greywall --learning -- opencode

# Block dangerous commands
greywall -c "rm -rf /"  # → blocked by command deny rules

Homebrew (macOS):

brew tap greyhavenhq/tap
brew install greywall

This also installs greyproxy as a dependency.

Linux / Mac:

curl -fsSL https://raw.githubusercontent.com/GreyhavenHQ/greywall/main/install.sh | sh

go install github.com/GreyhavenHQ/greywall/cmd/greywall@latest

mise use -g github:GreyhavenHQ/greywall
mise use -g github:GreyhavenHQ/greyproxy

git clone https://github.com/GreyhavenHQ/greywall
cd greywall
make setup && make build

Linux dependencies:

• bubblewrap - container-free sandboxing (required)
• socat - network bridging (required)
• xdg-dbus-proxy - filtered D-Bus proxy for notify-send support (optional)
• libsecret-tools - keyring credential injection for gh/glab (optional)

Check dependency status with greywall check.

# Run with all network blocked (default)
greywall -- curl https://example.com

# Run with shell expansion
greywall -c "echo hello && ls"

# Route through a SOCKS5 proxy
greywall --proxy socks5://localhost:1080 -- npm install

# Expose a port for inbound connections (e.g., dev servers)
greywall -p 3000 -c "npm run dev"

# Enable debug logging
greywall -d -- curl https://example.com

# Monitor sandbox violations
greywall -m -- npm install

# Show available Linux security features
greywall --linux-features

# Show version
greywall --version

# Check dependencies, security features, and greyproxy status
greywall check

# Install and start greyproxy
greywall setup

Greywall ships with built-in sandbox profiles for popular AI coding agents (Claude Code, Codex, Cursor, Aider, Goose, Gemini CLI, OpenCode, Amp, Cline, Copilot, Kilo, Auggie, Droid) and toolchains (Node, Python, Go, Rust, Java, Ruby, Docker).

On first run, greywall shows what the profile allows and lets you apply, edit, or skip:

$ greywall -- claude

[greywall] Running claude in a sandbox.
A built-in profile is available. Without it, only the current directory is accessible.

Allow read:  ~/.claude  ~/.claude.json  ~/.config/claude  ~/.local/share/claude  ~/.gitconfig  ...  + working dir
Allow write: ~/.claude  ~/.claude.json  ~/.cache/claude  ~/.config/claude  ...  + working dir
Deny read:   ~/.ssh/id_*  ~/.gnupg/**  .env  .env.*
Deny write:  ~/.bashrc  ~/.zshrc  ~/.ssh  ~/.gnupg

[Y] Use profile (recommended)   [e] Edit first   [s] Skip (restrictive)   [n] Don't ask again
>

Combine agent and toolchain profiles with --profile:

# Agent + Python toolchain (allows access to ~/.cache/uv, ~/.local/pipx, etc.)
greywall --profile claude,python -- claude

# Agent + multiple toolchains
greywall --profile opencode,node,go -- opencode

# List all available and saved profiles
greywall profiles list

Greywall can trace a command's filesystem access and generate a config profile automatically:

# Run in learning mode - traces file access via strace
greywall --learning -- opencode

# List generated profiles
greywall profiles list

# Show a profile's content
greywall profiles show opencode

# Next run auto-loads the learned profile
greywall -- opencode

Greywall reads from ~/.config/greywall/greywall.json by default (or ~/Library/Application Support/greywall/greywall.json on macOS).

{
  // Route traffic through an external SOCKS5 proxy
  "network": {
    "proxyUrl": "socks5://localhost:1080",
    "dnsAddr": "localhost:5353"
  },
  // Control filesystem access
  "filesystem": {
    "defaultDenyRead": true,
    "allowRead": ["~/.config/myapp"],
    "allowWrite": ["."],
    "denyWrite": ["~/.ssh/**"],
    "denyRead": ["~/.ssh/id_*", ".env"]
  },
  // Block dangerous commands
  "command": {
    "deny": ["git push", "npm publish"]
  }
}

Use greywall --settings ./custom.json to specify a different config file.

By default, traffic routes through the GreyProxy SOCKS5 proxy at localhost:43052 with DNS via localhost:43053.

See platform support for more details.

Greywall can also be used as a Go package.

Full documentation is available at https://docs.greywall.io/greywall.

• Quickstart Guide
• Why Greywall
• Configuration Reference
• Learning Mode
• Security Model
• Architecture
• Platform Support
• Linux Security Features
• AI Agent Integration
• Library Usage (Go)
• Troubleshooting

Greywall is a fork of Fence, originally created by JY Tan at Tusk AI, Inc. Copyright 2025 Tusk AI, Inc. Licensed under the Apache License 2.0.

Inspired by Anthropic's sandbox-runtime.