Class 13 · 13.4

Third-party cloud sandboxes: Sprites.dev

The newest, most pioneering way: don’t use Anthropic’s cloud at all — rent a specialist sandbox from a third party. We use Sprites.dev (by the fly.io team): a persistent Linux box that boots in under a second and runs any coding agent in full YOLO, because nothing it does can touch you.

🎯 Goal: run an agent in a rented cloud box⏱ Read: 15 min🤖 Type: optional / hands-on

This third approach is optional — you may be perfectly happy with the native sandbox and Anthropic’s cloud. But it is worth seeing, because it is where things are heading, and because it works with any agent, not just Claude.

01What a third-party sandbox is

Instead of Anthropic hosting the box, another company gives you a cloud container that is hardware-isolated and tuned for coding agents. Sprites.dev describes itself as a “stateful sandbox environment with checkpoint and restore” — a persistent Linux computer that is the simplest answer to “where should I run this blob of code?”

fast

Under a second

fly.io are proud of boot speed — sprite create returns in ~0.6s. No waiting around installing a server.

stateful

A home away from home

Checkpoint & restore: come back later and it is exactly as you left it. Your remote dev box, always ready.

isolated

Truly sandboxed

Hardware-isolated. Run in full YOLO — it can do whatever it wants and never reach your machine.

Create an account at Sprites.dev

They were offering ~$30 of trial credit (“500 sprites free”). It does ask for a credit card even with the free credits — entirely your choice whether to proceed or just watch.

Install the Sprite CLI

The first screen shows a one-line install command containing your key. Paste it into any terminal (your existing project terminal is fine).

Log in

sprite login pops a browser to sign in with the same account. Now the sprite command is ready.

Heads-up on the card

Sign-up bounces you to fly.io’s page (Sprites is built by fly.io) — just navigate back to Sprites.dev and you will be logged in. And never share that install command: it carries your key.

03Create a Sprite & clone your repo

Here is where it gets almost silly how easy it is:

terminal

sprite create finally-worker
Created finally-worker in 0.6 seconds.

# the prompt changes to  sprite@sprite  — you are now ON the remote box
sprite@sprite:~$ ls              # empty — a fresh cloud machine
sprite@sprite:~$ git clone <your finally repo>
sprite@sprite:~$ cd finally && ls
backend  CLAUDE.md  LICENSE  planning  README.md

That prompt matters

sprite@sprite tells you every command now runs on the remote box in the cloud — even though it looks just like your local terminal (especially inside VS Code’s terminal).

04Run Claude in safe YOLO

Claude Code comes pre-installed on every Sprite. Just type claude.

sprite@sprite

sprite@sprite:~/finally$ claude
# first-run questions (dark mode, etc.), then it wants to log in.
# Press 1 — it opens the browser on YOUR laptop, you authorise, done.

# It boots in "bypass permissions" (YOLO) mode automatically:
⚠ Claude Code running in bypass-permissions mode. Yes, I accept.
# Safe — it’s an isolated cloud box with a public repo. Nothing can hurt you.

A neat pro-tip

A fresh box may run a slightly older Claude Code (say, on an older model). Just quit and relaunch claude — it auto-updates itself to the latest version and model.

Now give it real work — it all runs remotely while your laptop is idle:

claude on the Sprite

> Read all docs in planning/. The market-data backend is implemented
  with tests. Carry out a comprehensive code review, run all the tests,
  and write your conclusions to planning/market-data-review.md

05The full remote workflow

The loop below is the whole magic — a remote box you drive like a local one:

You type in a terminal; the work runs on the remote box; results land in GitHub. It only feels local.

Review the run

Ctrl+O shows the full conversation — e.g. most tests passed, some failed, and Claude accounted for it.

Log in to GitHub from the box

Pushing needs gh auth login. Even though the box is remote, it pops the auth window on your laptop. After login, type claude then /resume to pick up exactly where you were.

Let it push & open a PR

“I’ve logged you in — try now.” It makes a branch, pushes, and opens a PR. Review & merge on GitHub.

Keep going, hands-off

“Now switch to main, pull, fix everything in the review file until all tests pass, then push a new branch.” It churns away — no approvals — and returns another PR.

The feeling is the best of both worlds: hugely productive (no permission nagging) and completely secure (fully sandboxed). That combination is what makes this so pleasant.

06It works for any agent

One last point: a Sprite does not care which agent you run. Claude Code, Codex, OpenCode — whichever you picked, it is ideal for running in a sandbox. That platform-independence is exactly why third-party sandboxes are catching on.

✓ Key takeaways

A third-party sandbox is a rented, hardware-isolated cloud box built for coding agents — e.g. Sprites.dev.
It is fast (<1s to create), stateful (resume where you left off) and isolated (safe full-YOLO).
Flow: sprite create → git clone → claude (pre-installed) → work → gh auth login → push → PR.
Login and auth windows still pop on your laptop, even though the box is remote.
It is agent-agnostic — run Claude, Codex or OpenCode inside it.