Recent submissions

Comment · 3 months ago
10 months ago I wanted to make a cloud coding agent software factory type of thing, a sort of Devin clone but more to my taste.

I started shopping for sandboxes and I realized like web hosting primitives in the web era, they were adding like a 50x markup on what qemu/docker + bare metal prices would be if I was doing it myself, even if I wanted full, beefy ubuntu VMs with all access.

Also generally their devX wasn't great, stuck between the old world of short-lived code execution for agents and now long-running agents. I struggled with fly a lot, struggled with e2b/daytona too, the daytona team even offered me 10k in credits when I started posting about my project but I knew long term I'd save hundreds of thousands rolling my own...

... and as I'm just a new grad with no savings but a lot of time, so I did it.

For 10 months I built 2 agentic factories, getting to a modest 1k MRR (BYOK, not selling tokens), with enough heavy power users to help figure out and fix a ton of edge cases that come with such builds.

At some point I ended up with a great sandbox underneath and decided to release it under a simple name: box

I wanted to make the simple, affordable sandbox for short to (very) long running agents I dreamed I had when building my own products.

So here we are, I launched 1.5m ago and it's getting quite popular, with hundreds of days of VM time consumed by our users per day, and real products with real paying users built by box users on top.

It comes as a CLI or API/SDKs. To create a new sandbox is just 8 keystrokes on the terminal 'box new' and each sandbox is a full ubuntu VM with a desktop (moonlight + VNC), all devtools, docker and hosting primitives pre-installed...

... and I've been able to offer it for like 5x to 25x less of a cost than other providers, I probably save some of my users thousands per week! And I still make a sustainable margin to pay myself and the infra, invest into a fast growth and development of box.

you can see a comparison with other sandboxes here: https://box.ascii.dev/compare and our docs at https://docs.ascii.dev/box

would love your take on sandboxes in general! it's such a fascinating space with so many interesting challenges and engineering compromises

Comment · 11 months ago
you mean you actually explore the code and all does it test it works?
Comment · 11 months ago
i don't get all these startups these days who just wrap oai and call it a day

cool ui tho

Comment · 11 months ago
Wanted to share this dependency hell post mortem, to remind everyone it's never easy to fully be safe from global outages, and if you want to be fail proof you really need to mind your dependencies and failure points on top of self-hosting a lot...
Comment · 1 year ago
That's very cool! I actually built a Claude Code Web alternative* over the last few months and made my own auto-commit solution for it. I reckon though having the prompt is what helps me and the agent generate relevant prompt messages that can explain the why. Anyway even a average commit message is way better than none

*ariana.dev

Comment · 1 year ago
Yeah that's why we basically built our own Claude Code Web but around Hetzner VPSs instead & terminal access. So you can use docker, open ports if you'd like. Some teams even needed us for a complicated R dev setup they wanted Claude to work with.
Comment · 1 year ago
Comment · 1 year ago
Hey HN!

We're two european builders who spent the last 2 months building Ariana, essentially Claude Code Web on steroids with some aspects ressembling Conductor.

We were frustrated with how managing multiple coding agents, sync and async, was feeling. Always either too much friction to go all-in most times, and once all-in too hard to manage. Especially as a team.

In Ariana you can launch multiple agents powered by your Claude Code on the same codebase, they go on to work on their own Hetzner VPS, on their own branch. They auto-commit, and you can generate a PR in the end.

You can launch & use agents from web, mobile web, desktop (win, mac, linux).

Why VMs over local git worktree? Well it doesn't clutter your disk space. It also makes agents shareable (one link share with your team), and useable from anywhere even away from the desktop. For example, you can start an agent on your local code and keep prompting it later in the metro.

Coding in a cloud VM is often annoying because you'd like to run or edit the code locally in your IDE. So on desktop we added real-time bidirectional file sync and network forwarding which allows you to fully focus on one agent as if it was there locally while the others keep working. You can also fork agents and by switching focus test multiple variants or features.

VMs can run anything like docker, docker compose, most languages. They each have an IP with firewall on by default. You can open ports and host what your agent worked on even.

We leave it for you to discover many other features and quality of life UX things.

You can try it on the web immediately for free using you claude subscription or API key. Linking a GitHub acc is optional (can just give a public clone url from github, gitlab, azure or bitbucket), but usage is more limited.

https://ariana.dev/app

If you want to try also the desktop, download it at: https://ariana.dev

Mixing desktop & web rhather than web only is highly recommended.

Thank you for your attention, curious to know what you think!

Live from the official Hacker News API

A modern reader for Hacker News · SvelteKit + Skeleton