Quick take: Sandboxed is the open-source backend that powers AI “build-in-a-browser” apps. One ./install.sh and you get isolated sandboxes, built-in coding agents, and auto preview URLs — no Kubernetes, no message queues, no SaaS dependencies.

But ever wanted to spin up something like Lovable or Bolt on your own server? And not just the frontend — the actual backend that creates sandboxes, runs AI agents, and serves live preview URLs.

I’ve been watching sandboxed (tastyeffectco/sandboxes) since it dropped 3 days ago. And 448 stars in 72 hours. I think it’s going to follow a curve similar to what we saw with Odysseus in its first week — except sandboxed fills a different gap.

So let me show you what I found after testing it.

What Sandboxed Actually Is

It’s a Go binary + Docker that gives every user their own isolated cloud dev environment. Plus built-in coding agents. Auto-assigned preview URLs. Sleep-wake cycles so idle sandboxes don’t eat your RAM.

And the whole architecture is deliberately boring:

No K8s. No message queue. No separate database. One Docker host handles it.

Sandboxed Install — One Command

So I ran this on a $6 VPS. It took under a minute:

git clone https://github.com/tastyeffectco/sandboxes.git
cd sandboxes
./install.sh

The script checks Docker, writes a .env, builds the images, and starts everything. Then curl http://127.0.0.1:9090/healthz returns ok.

And that’s it. You can start creating sandboxes and dispatching agents immediately.

Why Sandboxed Matters

Here’s the thing: right now, if you want to build an AI “code in the browser” product, you’re looking at months of infrastructure work — multi-tenant isolation, agent orchestration, preview URLs, crash recovery, auto-scaling. Or you pay for a SaaS like Lovable or Bolt and hand them your entire architecture.

But Sandboxed MIT-licenses that whole backend. Here’s how it stacks up:

Then I tested creating a sandbox and dispatching an OpenCode agent to build a Vite todo app. The agent cloned the repo, set up the dev server on port 3000, and sandboxed auto-assigned a s-<id>-3000.preview.localhost URL. From ./install.sh to a working preview in about 4 minutes.

One thing I noticed: the sleep-wake feature works, but the wake-up on the first request takes 2-3 seconds. Still, not a dealbreaker — way better than paying for idle sandboxes to stay hot all day. But on a $6 VPS with 2GB RAM, I managed 3 concurrent sandboxes before things got tight. A $12/mo Droplet would handle 8-10 comfortably.

I had a similar experience testing Headroom last week — another self-hosted tool that barely broke a sweat on modest hardware.

Who Sandboxed Is For

• Devs building AI coding products: Skip the infrastructure months
• Team leads who want data control: Your data, your server, your model keys
• Hobbyists who want their own Lovable: One server runs the whole thing
• Anyone tired of SaaS pricing on app-builder platforms

That said, not for non-technical folks yet — you still need Docker and a VPS. But the docs are clean and the codebase is readable in an afternoon.

Sandboxed: The Bottom Line

Honestly, Sandboxed scores an 8.08/10 in my book. It’s early (3 days, 5 open issues), but the core experience is solid — I went from git clone to a working AI sandbox in under 5 minutes. And for a project that’s existed less than a week, that’s impressive.

Plus, I’m running it on my test server and plan a full deployment guide soon. If you want to try it now:

Disclosure: Some links below are affiliate links. If you purchase through these links I may earn a small commission at no extra cost to you.

🔥 Ready to run Sandboxed?

Sandboxed runs on any Docker-capable VPS. I tested it on a $6 Droplet — here's what I recommend:

• DigitalOcean — Deploy a Docker droplet in 60 seconds. Get $200 free credit →
• Vultr — High-performance cloud VPS with global locations. Start with $50-100 free credit →

Want to try Sandboxed but don’t have a server? New DigitalOcean users get $200 in credit — that’s enough to run a $6/month Droplet for over two years. Or grab a Vultr $50-100 credit if you prefer their network. Both work great for Docker-based deployment.