How to sandbox Claude Code with nono

Claude Code is one of the most capable coding agents available today. It can navigate your codebase, edit files across multiple directories, run shell commands, install packages, and execute tests — all autonomously.

It also runs with your full user permissions. Every file on your system is readable. Every credential in your environment is accessible. Every command it runs executes as you.

We've all seen the stories. Claude deleted my database. Claude deleted all my code. Claude wiped my home directory. These aren't hypothetical risks — they're things that happen when an agent operates with no isolation between what it wants to do and what it can do.

Permission prompts don't scale

The default response to this problem has been permission dialogs. The agent wants to run a command — do you approve? The agent wants to write a file — do you approve?

The issue is predictable. After the third prompt, you're clicking "allow" reflexively. Approval fatigue turns a security feature into a rubber stamp. You end up with interrupted flow and reduced security.

Claude Code now ships with a built-in sandbox mode, which is a meaningful step forward. It uses OS-level primitives to enforce filesystem and network boundaries, reducing the need for constant approval prompts.

But it includes an escape hatch. When a command fails due to sandbox restrictions, Claude can retry it outside the sandbox with the dangerouslyDisableSandbox parameter. The name is honest — it is dangerous. One retry outside the sandbox and the agent has full access again.

Kernel-level enforcement with no escape

nono takes a different approach. It uses Landlock (Linux) and Seatbelt (macOS) to enforce restrictions at the kernel level. Once applied, these restrictions are irreversible for the process. There is no API to widen them. There is no escape hatch. Not even nono itself can remove the restrictions after they're set.

This is what we mean by "structurally impossible." The kernel denies unauthorised operations directly. There's no interception layer, no policy filter, no mechanism to bypass. If the agent tries to read your SSH keys or delete files outside its granted path, the operation fails at the syscall level.

Live demo: sandboxing Claude Code

We recorded a short demo showing exactly what this looks like in practice.

In the video, we install nono via Homebrew, launch Claude Code inside a nono sandbox using the built-in Claude profile, and then attempt to trick the agent into doing things it shouldn't:

Reading AWS credentials — we ask Claude to access a confidential file used for AWS account access. The operation is immediately denied. Claude is aware it's running inside a nono sandbox and tells the user how to explicitly grant access if needed.

Deleting a database — we start with an innocent-sounding question, then escalate to asking Claude to delete a database. Claude attempts to access a directory outside its granted boundary. The kernel blocks it. Claude helpfully explains what happened and how to widen permissions if the user genuinely wants to.

The whole setup takes about 30 seconds. Install via Homebrew, wrap your Claude Code session with nono, done.

Getting started

Install nono:

brew tap always-further/nono
brew install nono

Launch Claude Code inside a sandbox using the built-in profile:

nono run --allow-cwd --profile claude-code -- claude

That's it. Claude Code now runs with default-deny filesystem access. Only the current directory is writable. SSH keys, AWS credentials, system configs — all blocked at the kernel level.

If you want to run silently without the nono banner:

nono run --allow-cwd --silent --profile claude-code -- claude

For more granular control:

# Read-only access to reference docs, write access to project
nono run --read ~/docs --write ./src --profile claude-code -- claude

# Check why a specific path would be blocked
nono why --path ~/.ssh/id_rsa --op read

What's next

nono is free, open source, and available today. We're actively developing fine-grained network controls, atomic rollbacks, and library bindings for Python, TypeScript, and Rust so other tools can integrate nono's security primitives directly.

• Docs: docs.nono.sh
• GitHub: github.com/always-further/nono
• Discord: Join the community

Don't yolo when you can nono.