agents / claude

Stonewall for Claude Code

Stonewall runs Claude Code inside a kernel-enforced sandbox. The agent only sees your project directory and the few binaries and paths the policy names. Everything else on your machine, from ~/.ssh to your browser profile, is invisible to it, no matter what a prompt says.

Policy

For Claude Code, stonewall.sh provides a specific policy at https://stonewall.sh/policies/claude.yml. You can use it by simply including it in your .stonewall.yml config, or using the CLI:

stonewall policy include https://stonewall.sh/policies/claude.yml

It covers only what Claude needs to start and authenticate:

• Binaries: claude, node (the runtime) and security (macOS keychain, used for credentials).
• Writable paths: ~/.claude and ~/.claude.json, where sessions, settings and login state live.

It intentionally does not allow sh, bash, git or any other tool the agent could use to run arbitrary programs.

Related Policies

The Claude Code policy is meant to be combined with policies for what your project actually needs:

• base.yml: read-only file inspection (cat, grep, ls, …), hides .env and secrets.
• git-safe.yml: read-only git, no commits, no credentials.
• git-unsafe.yml: full git, including commits and push.
• A language policy for builds and tests, e.g. go.yml, python.yml or typescript.yml.

Usage

Start the agent through stonewall instead of directly:

stonewall claude

Or alias it in your bash / zsh profile:

alias claude="stonewall claude"