Class 13 · 13.5

Sandboxing recap & cheat-sheet

One page to rule them all. Here is the whole of Class 13 — the three approaches side by side, when to reach for each, every command in one place, and a short self-check so you know it stuck.

🎯 Goal: consolidate & keep as reference⏱ Read: 10 min📝 Type: reference

Today we made YOLO safe three different ways. If any of it blurred together, this page is your anchor — bookmark it.

01The three approaches, one picture

Blue runs on your laptop; purple and yellow run in the cloud. All three let Claude work freely without touching anything it shouldn’t.

02Which one should I use?

If you want to…	Use	Why
Auto-approve safe work on your own laptop	Blue · /sandbox	Lightest, instant, no accounts. (Windows: needs WSL.)
Fire off tasks and keep your laptop free	Purple · web	Runs on Anthropic’s machines; you review a PR.
Work from your phone / on the move	Purple · mobile	Same cloud sessions, from the Claude app.
Clear a backlog in parallel	Purple · @claude	Tag many GitHub issues; a fleet of cloud Claudes.
A fast, persistent remote dev box for any agent	Yellow · Sprites	Isolated, stateful, agent-agnostic, full-YOLO safe.

03Command cheat-sheet

every command from today

# --- BLUE: native sandbox ---
/sandbox                     # configure & enable (pick mode 1 = sandboxed YOLO)

# --- PURPLE: managed cloud (one-time setup) ---
/install-github-app          # installs gh CLI + the two workflows
gh auth login                 # log this machine into GitHub
# then, five ways to run remote:
> & <your task>               # 1. ampersand prefix, inside Claude
claude --remote "<task>"       # 2. from the terminal
# 3. browser  claude.ai/code   4. Claude mobile app  Code tab
# 5. GitHub issue, tag  @claude
/tasks                       # see everything running remotely

# --- YELLOW: Sprites.dev ---
<paste install one-liner>     # installs sprite CLI (contains your key)
sprite login
sprite create my-worker       # <1s; prompt becomes sprite@sprite
git clone <repo> && cd <repo>
claude                        # pre-installed; /resume to continue a session

04Security checklist

Before you go full-YOLO

Keep explicit deny rules in your sandbox config — they are always respected.
In the native sandbox, allow-list only the sites you actually need for the network.
Never paste or share a command that contains a key (the Sprite install line does).
Full-YOLO is safe because the box is isolated — only run it inside a real sandbox, never on your bare machine.
Read the sandbox docs’ security section once if you will use it heavily.

The human still owns the merge

Cloud and sandbox make Claude fast and safe to run. They do not review the code for you. Every branch still comes back as a PR that you read and merge.

05Self-check

Why is un-sandboxed “safe” sometimes riskier?

Approval fatigue — you rubber-stamp prompts on autopilot instead of reading them.

Which approach doesn’t work on plain Windows yet?

The native /sandbox — it needs WSL for now.

What travels to the cloud with an `&` task?

The full conversation context, so the remote Claude continues seamlessly.

Why is full-YOLO safe on a Sprite?

It’s a hardware-isolated cloud box — nothing it does can reach your machine.

✓ Key takeaways

Three ways to make YOLO safe: Blue (local /sandbox), Purple (managed cloud), Yellow (Sprites).
Pick by where you want the work to run and whether you need parallelism or mobility.
Keep deny-rules, allow-list narrowly, guard your keys — and always review the PR.
You are 80% through the course. Tomorrow: big codebases, then the frontier — the Agent SDK, Cowork and OpenClaw.