YOLO Mode in a Vagrant Box

A Vagrant-based sandbox for running AI coding agents in full autonomous mode—without risking your host machine.

The problem #

AI coding agents work best when they can run autonomously. Claude Code’s claude --dangerously-skip-permissions and Codex’s codex --full-auto let agents execute commands without asking for approval at each step. But this is scary for good reason—agents have deleted home directories, wiped working directories, and destroyed files while misreporting what happened.

Solution: disposable VMs #

People use devcontainers, Docker, or remote VMs to isolate their agents. I wanted something I can blow away and recreate in seconds. Vagrant is a bit of an overkill, but it works.

Note the trust boundary: the VM gives you OS-level isolation, but your repo inside is fully writable—if the agent deletes your code, it’s gone (until you vagrant destroy and start fresh from git). The setup also forwards your SSH agent and mounts git config, so treat the VM as having your git credentials.

Quick start #

# Clone the sandbox repo
git clone https://github.com/igor-kupczynski/yolo-sandbox.git
cd yolo-sandbox

# Start the VM and SSH in
vagrant up && vagrant ssh

# Inside the VM: clone your project and run the agent
git clone <your-repo>
cd <your-repo>
claude --dangerously-skip-permissions  # or: codex --full-auto

Try it: yolo-sandbox. Inspired by Emil Burzo’s post.