██╗   ██╗ ██████╗ ██╗      ██████╗ ██████╗  ██████╗ ██╗  ██╗
╚██╗ ██╔╝██╔═══██╗██║     ██╔═══██╗██╔══██╗██╔═══██╗╚██╗██╔╝
 ╚████╔╝ ██║   ██║██║     ██║   ██║██████╔╝██║   ██║ ╚███╔╝
  ╚██╔╝  ██║   ██║██║     ██║   ██║██╔══██╗██║   ██║ ██╔██╗
   ██║   ╚██████╔╝███████╗╚██████╔╝██████╔╝╚██████╔╝██╔╝ ██╗
   ╚═╝    ╚═════╝ ╚══════╝ ╚═════╝ ╚═════╝  ╚═════╝ ╚═╝  ╚═╝

Let your AI go full send. Your home directory stays home.

Docs: yolobox.dev

Run Claude Code, Codex, or any AI coding agent in "yolo mode" without nuking your home directory.

The Problem

AI coding agents are incredibly powerful when you let them run commands without asking permission. But one misinterpreted prompt and rm -rf ~ later, you're restoring from backup (yea right, as if you have backups lol).

The Solution

yolobox runs your AI agent inside a container where:

• ✅ Your project directory is mounted at its real path (e.g., /Users/you/project)
• ✅ The agent has full permissions and sudo inside the container
• ✅ Your home directory is NOT mounted (unless you explicitly opt in)
• ✅ Persistent volumes keep tools and configs across sessions
• ✅ Session continuity - AI sessions can be resumed across host/container boundary

The AI can go absolutely wild inside the sandbox. Your actual home directory? Untouchable.

Quick Start

# Install via Homebrew (macOS/Linux)
brew install finbarr/tap/yolobox

# Or install via script (requires Go)
curl -fsSL https://raw.githubusercontent.com/finbarr/yolobox/master/install.sh | bash

Then from any project:

cd /path/to/your/project
yolobox claude    # Let it rip

Or use any other AI tool: yolobox codex, yolobox gemini, yolobox copilot.

Non-interactive invocations keep stdout and stderr separate, so shell redirection works as expected:

yolobox claude -- -p "Hello" 2>/dev/null

What's in the Box?

The base image comes batteries-included:

• AI CLIs: Claude Code, Gemini CLI, OpenAI Codex, OpenCode, Copilot (all pre-configured for full-auto mode!)
• Runtimes: Node.js 22, Python 3, Go, Bun
• Build tools: make, cmake, gcc
• Git + GitHub CLI
• Common utilities: ripgrep, fd, fzf, jq, vim

Need something else? The AI has sudo.

AI CLIs Run in YOLO Mode

Inside yolobox, the AI CLIs are aliased to skip all permission prompts:

Command	Expands to
claude	claude --dangerously-skip-permissions
codex	codex --ask-for-approval never --sandbox danger-full-access
gemini	gemini --yolo
opencode	opencode (no yolo flag available yet)
copilot	copilot --yolo

No confirmations, no guardrails—just pure unfiltered AI, the way nature intended.

For Codex, yolobox pins approval and sandbox mode explicitly so upstream trust defaults and Linux sandbox backend changes do not change the wrapper behavior.

Project-Level Container Customization

Need extra tools for one project without bloating the base image? yolobox can build and cache a derived image from your project config.

Simple case:

# .yolobox.toml
[customize]
packages = ["default-jdk", "maven"]

Then run normally:

yolobox run mvn --version

For more advanced setups, add a Dockerfile fragment:

# .yolobox.toml
[customize]
dockerfile = ".yolobox.Dockerfile"

# .yolobox.Dockerfile
USER root
RUN curl -fsSL https://get.sdkman.io | bash
USER yolo

You can combine both. packages install first, then the fragment runs on top.

For one-offs, use flags:

yolobox run --packages default-jdk,maven mvn --version
yolobox run --customize-file .yolobox.Dockerfile bash
yolobox run --packages default-jdk --rebuild-image java --version

The first run builds a derived image. Later runs reuse it until the base image or customization inputs change. When you use a Dockerfile fragment, yolobox asks Docker/Podman to build again so context changes are noticed, but cached layers are reused when nothing changed.

This also keeps yolobox upgrade relatively painless:

• yolobox upgrade still updates the binary and pulls the latest base image
• your project-level customization stays in .yolobox.toml / .yolobox.Dockerfile
• the next run rebuilds the derived image only if the new base image or your customization inputs changed

You still pay one rebuild after a base-image upgrade, but you do not need to manually rebase a forked Dockerfile just to keep using the feature.

Note: Derived-image customization requires a runtime that can build images (docker or podman). Apple's container runtime can run yolobox, but it cannot build custom images.

Runtime Support

• macOS: Docker Desktop, OrbStack, Colima, or Apple container (macOS Tahoe+)
• Linux: Docker or Podman

yolobox auto-detects available runtimes. To use a specific runtime:

yolobox claude --runtime container   # Apple container
yolobox claude --runtime docker      # Docker
yolobox claude --runtime podman      # Podman

Memory: Claude Code needs 4GB+ RAM allocated to Docker. Colima defaults to 2GB which will cause OOM kills. Increase with: colima stop && colima start --memory 8

Commands

# AI tool shortcuts (recommended)
yolobox claude              # Run Claude Code
yolobox codex               # Run OpenAI Codex
yolobox gemini              # Run Gemini CLI
yolobox opencode            # Run OpenCode
yolobox copilot             # Run GitHub Copilot

# General commands
yolobox                     # Drop into interactive shell (for manual use)
yolobox run <cmd...>        # Run any command in sandbox
yolobox setup               # Configure yolobox settings
yolobox upgrade             # Update binary and pull latest image
yolobox config              # Show resolved configuration
yolobox reset --force       # Delete volumes (fresh start)
yolobox version             # Show version
yolobox help                # Show help

Configuration

Run yolobox setup to configure your preferences with an interactive wizard.

Settings are saved to ~/.config/yolobox/config.toml:

git_config = true
gh_token = true
ssh_agent = true
docker = true
no_network = true
network = "my_compose_network"
no_yolo = true
cpus = "4"
memory = "8g"
cap_add = ["SYS_PTRACE"]
devices = ["/dev/kvm:/dev/kvm"]
runtime_args = ["--security-opt", "seccomp=unconfined"]

You can also create .yolobox.toml in your project for project-specific settings:

mounts = ["../shared-libs:/libs:ro"]
env = ["DEBUG=1"]
readonly_project = true
exclude = [".env*", "secrets/**"]
copy_as = [".env.sandbox:.env"]
no_network = true
shm_size = "2g"

Priority: CLI flags > project config > global config > defaults.

Each runtime_args entry is a single CLI argument. For flags that take a value, add them as separate entries so --security-opt seccomp=unconfined becomes ["--security-opt", "seccomp=unconfined"].

Note: Setting claude_config = true or gemini_config = true in your config will copy your host config on every container start, overwriting any changes made inside the container (including auth and history). Prefer using --claude-config or --gemini-config for one-time syncs.

Project File Filtering

Use --exclude to hide matching project paths from the container by staging an empty readonly project view:

yolobox claude --readonly-project --exclude ".env*" --exclude "secrets/**"

Use --copy-as to copy one file into another project path inside that staged readonly view:

yolobox claude --readonly-project --exclude ".env*" --copy-as ".env.sandbox:.env"

• exclude patterns are relative to the project root and support **
• copy_as destinations must stay inside the project and must already exist as files
• copy_as wins if it targets the same path as an exclude
• --exclude and --copy-as currently require --readonly-project
• --exclude and --copy-as are currently supported on Docker and Podman, not Apple's container runtime

Copying Global Agent Instructions

The --copy-agent-instructions flag copies your global/user-level agent instruction files into the container. This is useful when you have custom rules or preferences defined globally that you want available inside yolobox.

Files copied (if they exist on your host):

Tool	Source	Destination
Claude	~/.claude/CLAUDE.md	/home/yolo/.claude/CLAUDE.md
Gemini	~/.gemini/GEMINI.md	/home/yolo/.gemini/GEMINI.md
Codex	~/.codex/AGENTS.md	/home/yolo/.codex/AGENTS.md
Copilot	~/.copilot/agents/	/home/yolo/.copilot/agents/

Note: This only copies global instruction files, not full configs (credentials, settings, history). For Claude's full config, use --claude-config instead.

You can also set copy_agent_instructions = true in your config file for persistent use.

Auto-Forwarded Environment Variables

These are automatically passed into the container if set:

• ANTHROPIC_API_KEY
• CLAUDE_CODE_OAUTH_TOKEN
• OPENAI_API_KEY
• COPILOT_GITHUB_TOKEN / GH_TOKEN / GITHUB_TOKEN
• OPENROUTER_API_KEY
• GEMINI_API_KEY

Note: On macOS, gh CLI stores tokens in Keychain, not environment variables. Use --gh-token (or gh_token = true in config) to extract and forward your GitHub CLI token.

Flags

Note: Flags go after the subcommand: yolobox run --flag cmd or yolobox claude --flag, not yolobox --flag run cmd.

Flag	Description
--runtime <name>	Use docker, podman, or container (Apple)
--image <name>	Custom base image
--mount <src:dst>	Extra mount (repeatable)
--exclude <glob>	Hide matching project paths from the container (repeatable)
--copy-as <src:dst>	Mount a file at another project path inside the container (repeatable)
--env <KEY=val>	Set environment variable (repeatable)
--setup	Run interactive setup before starting
--ssh-agent	Forward SSH agent socket
--no-network	Disable network access
--network <name>	Join specific network (e.g., docker compose)
--pod <name>	Join existing Podman pod (shares its network)
--no-yolo	Disable auto-confirmations (mindful mode)
--scratch	Start with a fresh home/cache (nothing persists)
--readonly-project	Mount project read-only (outputs go to /output)
--claude-config	Copy host ~/.claude config into container
--gemini-config	Copy host ~/.gemini config into container
--git-config	Copy host ~/.gitconfig into container
--gh-token	Forward GitHub CLI token (extracts from keychain via gh auth token)
--copy-agent-instructions	Copy global agent instruction files (see configuration below)
--docker	Mount Docker socket and join shared network (see notes below)
--cpus <num>	Limit CPUs available to the container (accepts fractions like 3.5)
--memory <limit>	Hard memory limit (e.g., 8g, 1024m)
--shm-size <size>	Size of /dev/shm tmpfs (useful for browsers/playwright)
--gpus <spec>	Pass GPUs (Docker/Podman notation, e.g., all, device=0)
--device <src:dest>	Add host devices in the container (repeatable)
--cap-add <cap>	Add Linux capabilities (repeatable)
--cap-drop <cap>	Drop Linux capabilities (repeatable)
--runtime-arg <flag>	Pass raw runtime flags directly to Docker/Podman (repeatable)
--packages <list>	Comma-separated apt packages for a derived custom image
--customize-file <path>	Dockerfile fragment for a derived custom image
--rebuild-image	Force rebuild of the derived custom image

Resource & security controls: The table lists the common knobs baked into yolobox. Anything else (e.g., --ulimit nofile=4096:8192, --security-opt seccomp=unconfined) can be forwarded verbatim with --runtime-arg <flag> as many times as needed. Docker and Podman accept the passthrough flags unchanged; Apple's container runtime ignores options it doesn't understand.

SSH agent (macOS): On macOS, --ssh-agent requires the Docker VM to forward the SSH agent. For Colima: edit ~/.colima/default/colima.yaml, set forwardAgent: true, then restart (colima stop && colima start). Docker Desktop forwards the agent automatically.

Networking: By default, yolobox uses Docker's bridge network (internet access, no container DNS). Use --network <name> to join a docker compose network and access services by name. Use --no-network for complete isolation.

Docker access: The --docker flag mounts the host Docker socket into the container and joins a shared yolobox-net network. This lets the AI agent run Docker commands (build images, start containers, use docker compose) that create sibling containers on the same network. The agent and any services it creates can communicate by container name. The network name is available inside the container as $YOLOBOX_NETWORK. Cannot be used with --no-network.

Project filtering: --exclude globs are evaluated relative to the project root. --copy-as destinations must already exist as files in the project. Both flags currently require --readonly-project. Apple's container runtime does not support them yet.

Philosophy: It's the AI's Box, Not Yours

yolobox is designed for AI agents, not humans. The typical workflow is:

yolobox claude    # Launch Claude Code in the sandbox
yolobox codex     # Or Codex, or any other AI tool

That's it. You launch the AI and let it work. You're not meant to manually enter the box and set things up—the AI does that itself.

Why? The AI agent has full sudo access inside the container. If it needs a compiler, database, or framework—it just installs it. Named volumes persist these installations across sessions, so setup happens once. You point it at your project and let it cook.

Security Model

How It Works

yolobox uses container isolation (Docker or Podman) as its security boundary. When you run yolobox, it:

1. Starts a container with your project mounted at its real path
2. Runs as user yolo with sudo access inside the container
3. Does NOT mount your home directory (unless explicitly requested)
4. Uses Linux namespaces to isolate the container's filesystem, process tree, and network

The AI agent has full root access inside the container, but the container's view of the filesystem is restricted to what yolobox explicitly mounts.

Trust Boundary

The trust boundary is the container runtime (Docker/Podman). This means:

• ✅ Protection against accidental rm -rf ~ or credential theft
• ✅ Protection against most filesystem-based attacks
• ⚠️ NOT protection against container escapes — a sufficiently advanced exploit targeting kernel vulnerabilities could break out
• ⚠️ NOT protection against a malicious AI deliberately trying to escape — this is defense against accidents, not adversarial attacks

If you're worried about an AI actively trying to escape containment, you need VM-level isolation (see "Hardening Options" below).

Threat Model

What yolobox protects:

• Your home directory from accidental deletion
• Your SSH keys, credentials, and dotfiles
• Other projects on your machine
• Host system files and configurations

What yolobox does NOT protect:

• Your project directory (it's mounted read-write by default)
• Network access (use --no-network to disable, or --network <name> for specific networks)
• The container itself (the AI has root via sudo)
• Against kernel exploits or container escape vulnerabilities

If you want a narrower view of the project, use --exclude and --copy-as to hide or replace selected files before the agent sees them.

Hardening Options

Level 1: Basic (default)

yolobox  # Standard container isolation

Level 2: Reduced attack surface

yolobox claude --no-network --readonly-project --exclude ".env*" --exclude "secrets/**"

Level 3: Rootless Podman (recommended for security-conscious users)

# Install podman and run rootless
yolobox --runtime podman

Rootless Podman runs the container without root privileges on the host, using user namespaces. This significantly reduces the impact of container escapes since the container's "root" maps to your unprivileged user on the host.

Level 4: VM isolation (maximum security)

For true isolation with no shared kernel, consider running yolobox inside a VM:

• macOS: Use a Linux VM via UTM, Parallels, or Lima
• Linux: Use a Podman machine or dedicated VM

This adds significant overhead but eliminates kernel-level attack surface.

Network Isolation with Podman

For users who want to prevent container access to the local network while preserving internet access:

# Rootless podman uses slirp4netns by default, which provides
# network isolation from the host network
podman run --network=slirp4netns:allow_host_loopback=false ...

yolobox doesn't currently expose this as a flag, but you can achieve it by running rootless Podman (the default network mode for rootless is slirp4netns).

Building the Base Image

make image

This builds ghcr.io/finbarr/yolobox:latest locally, overriding the remote image.

Customizing the Image

Need more control than packages or a small Dockerfile fragment? You can still build and use a fully custom image:

1. Clone and modify:

git clone https://github.com/finbarr/yolobox.git
cd yolobox
# Edit the Dockerfile to add your packages

2. Build with a custom name:

make image IMAGE=my-yolobox:latest

3. Configure yolobox to use it:

mkdir -p ~/.config/yolobox
echo 'image = "my-yolobox:latest"' > ~/.config/yolobox/config.toml

Using a custom image name means yolobox upgrade won't overwrite your customization. That is the upside.

The downside is that you now own the drift:

• upstream Dockerfile changes do not automatically flow into your custom image
• yolobox upgrade will update the binary, but it will not rebuild or migrate your custom image for you
• when the base image changes upstream, you need to pull those changes into your fork and rebuild manually
• the farther your custom Dockerfile drifts, the more upgrade work you take on

If you mostly need "add a few tools for this project", prefer project-level customization above. Use a fully custom image only when you need full control over the entire base image.

Development

Building

make build          # Build binary
make test           # Run tests
make lint           # Run linters
make image          # Build Docker image
make install        # Install to ~/.local/bin

Versioning

Version is derived automatically from git tags via git describe:

• Tagged commit: v0.1.1
• After tag: v0.1.1-3-gead833b (3 commits after tag)
• Uncommitted changes: adds -dirty

No files to edit for releases. The Makefile handles it.

Releasing

To release a new version:

git tag v0.1.2
git push origin master --tags

That's it. GitHub Actions will automatically:

1. Build binaries for linux/darwin × amd64/arm64
2. Create a GitHub release with binaries and checksums
3. Build and push Docker image to ghcr.io/finbarr/yolobox

Version policy:

• Patch bump (0.1.x): Bug fixes, security fixes
• Minor bump (0.x.0): New features
• Major bump (x.0.0): Breaking changes

License

MIT