One server for every idea

In September I wrote that I was looking at moving my projects onto a single server. I’ve done it now. Everything I build lives in two places: a VPS from Hostup, a Swedish host, and Cloudflare.

Before, it was spread out. Hosting was on Vercel, databases and logins on Supabase, and domains at Loopia, where they had been for years. Now the domains and DNS are at Cloudflare, static sites like this one run on Cloudflare for free, and anything with a database runs on the VPS. It’s a cheap server but a good one, and it costs the same each month whether it runs two apps or twenty.

Coolify does the boring part

On the server I run Coolify, an open source, self-hosted take on what Vercel does. You connect a GitHub repo, and every push builds and deploys. Each project gets its own Postgres database next to it.

What I like is how little it takes to start something. A new idea gets a repo, a database, a subdomain and a deploy, and none of it adds a line to a bill. So I can build a first version that’s fairly advanced, with real logins, real data and real emails, put it on a real address and see how it feels to use. Most ideas don’t survive that, and it costs almost nothing to find out.

Claude Code runs the infrastructure

The part that makes this work for me is that I don’t operate any of it by hand. Claude Code can reach Coolify, Cloudflare and GitHub, so it creates the DNS records, sets up the project, adds the database and deploys. When something breaks, it reads the logs. I’m a designer and I don’t write code, and I still never need to log into a dashboard to start a project.

That only works if the AI knows how I want things done, so I made two things.

A rulebook and a starter

The first is an infra project, a single document with the rules every project follows. Where things are hosted, how secrets are handled, how a project is deployed and backed up, and how the server is locked down. Every project points to it, so Claude Code reads the same rules every time. When I learn something, like the mistakes I made moving this site, it goes in there once and every project after it benefits.

The second is a starter template. A new app starts with logins, a database, email, payments, file storage, analytics and the security settings already in place and wired to the rules. API keys and tokens have fixed names and a fixed place, and the AI isn’t allowed to read them. Starting something new means copying the template, switching off what the idea doesn’t need and describing what it should do.

It took a while to set up. Now the first hour of a new project goes into the idea instead of the plumbing.

Old ideas get a second chance

In the September post I wrote about Supabase’s free plan, which allows two projects. Twice I shut one down to make room for a new idea. Those projects didn’t die because nobody wanted them. They died because of a plan limit.

That limit is gone. Every project gets its own database on the same server at no extra cost, so I can go back to ideas I put away, or never properly tried, and give them a fair test this time.

Tools just for me

It also changes what’s worth building at all. A tool only I would use used to be hard to justify, with its own hosting, its own database and its own bill. Now it’s one more project on the same server. Appscope started that way, because I wanted one place to see how all my apps are doing. Vaka too: it watches my bank accounts and emails me only when something is worth knowing. They’re small, and they make my own work a bit easier.

I might even build my own AI assistant like this, running on my own server with access to my own tools and data.

The Lab

To keep track of it all, I added a Lab page to this site. It lists everything I’m building and testing on my own, including the tools only I use.

Why this suits building with AI

When AI writes the code, building is no longer the slow part. What slows you down is everything around it: accounts, dashboards, bills, setup, and the weight of one more thing to look after. When that part is shared, written down and run by the same AI that writes the code, trying an idea gets close to free.