The problem wasn’t the price
For years my sites sat on a traditional web host. It worked, but it was static and inflexible: one PHP version, one way of deploying, one control panel. If I wanted to test something, the hosting set the limit, not me. And every year I paid for capacity I didn’t use.
Saving money was a bonus. The real reason I left was that the host had become the slowest part of how I work.
What the setup looks like now
- Most of my sites run on Vercel. They’re static or framework-based, they deploy on push, and every branch gets its own preview URL. The host doesn’t need to understand PHP for any of it. It’s the same static-first approach that kodekid.app itself is built on.
- WordPress projects I’m developing run on a Linux machine in a broom cupboard. It’s a small used mini PC running Proxmox. Each project gets its own container with WordPress, MariaDB and a Cloudflare Tunnel in Docker. That’s where I build custom WordPress themes from scratch, with no hosting panel deciding what’s allowed.
- Mail moved to Migadu. Four addresses on my own domain for 127 DKK a year.
- DNS stayed at Cloudflare, which ties everything together.
The result is about 1,600 DKK a year saved on hosting and mail. The mini PC was already running Home Assistant for the house, so the extra power cost is close to zero. The bigger win is control.
One caveat before you copy this: Vercel’s free Hobby plan is for personal, non-commercial use only. Client work and anything that earns money belongs on Pro, so factor that into your own numbers.
How a client site runs from a cupboard without opening the router
A WordPress site I’m building for a client doesn’t need to be public. It needs to be reachable for me and the client while we’re working on it, and closed the rest of the time.
- Cloudflare Tunnel instead of port forwarding. The container opens an outbound-only connection to Cloudflare, so no ports are open on my router. It even works behind CGNAT.
- Cloudflare Access in front. You only reach the site with an email address on an allow list. You get a one-time code by mail, so there are no passwords to share and no accounts for the client to create. For a small team, Access is free.
- On when needed, off otherwise. A small script starts the container for a set time, say three hours, and shuts it down afterwards. The site isn’t online when nobody is using it.
- Snapshot before anything risky. Before a migration or a plugin update, I take a snapshot of the container. If something breaks, I’m back in seconds. My old host never offered anything close.
The on/off part doesn’t need to be clever. A minimal version is two commands on the Proxmox host:
pct start 101
systemd-run --on-active=3h /usr/sbin/pct shutdown 101
What went wrong (so you can skip it)
The move wasn’t flawless, and the mistakes are the useful part.
1. The wp-content copy was incomplete
The site looked fine on my desktop, but 40 files from one upload folder were missing. Logo, favicon and team photos only showed up because my browser had them cached.
Lesson: compare the files the media library references in the database with what’s actually on disk. Don’t trust the frontend. Run this from the WordPress root:
wp db query "SELECT meta_value FROM $(wp db prefix)postmeta WHERE meta_key = '_wp_attached_file'" --skip-column-names \
| while IFS= read -r f; do [ -f "wp-content/uploads/$f" ] || echo "MISSING: $f"; done
2. Old URLs survive the search-replace
Elementor writes its own CSS files with absolute URLs, and some content had URLs without https:. Those pointed to the old domain, which now redirects somewhere else, so images went missing on phones and in Safari.
Lesson: search for the old domain in the database and in wp-content, then regenerate the page builder’s CSS. Match on //old-domain rather than the full URL, so you catch http://, https:// and protocol-relative links without touching email addresses. Elementor also stores its layouts as JSON with escaped slashes, which a normal search-replace walks straight past, so it needs its own command:
wp search-replace '//old-domain.dk' '//new-domain.dk' --all-tables --dry-run
wp search-replace '//old-domain.dk' '//new-domain.dk' --all-tables
wp search-replace 'http://new-domain.dk' 'https://new-domain.dk' --all-tables
wp elementor replace-urls https://old-domain.dk https://new-domain.dk
wp elementor flush-css
grep -rIl 'old-domain.dk' wp-content/
3. “One-time PIN” isn’t always the default login
Once another login method has been added to Cloudflare Access, the email code has to be enabled explicitly on the application. My client landed on a Cloudflare sign-in page they couldn’t use.
Lesson: test the login in a private window with the client’s own address before you send the link.
4. A VPN client can quietly take over DNS
Tailscale had emptied the server’s DNS configuration, and the first install failed with name-resolution errors.
Lesson: if downloads fail on a fresh box, check /etc/resolv.conf first. Tailscale rewrites that file when MagicDNS is on and no DNS manager is running. tailscale set --accept-dns=false stops it on that machine.
Where it stops
This isn’t free hosting. It’s hosting where I’m the support desk.
- Operations and backups are mine. If the power goes in the cupboard, the site is down. That’s fine for development and staging, and not okay for a client’s live production site.
- A snapshot isn’t a backup. It lives on the same disk as the container. It saves you from a bad plugin update, not from a dead SSD. Real backups go to another machine, or at least another disk.
- Home internet isn’t a data centre. Upload speed, uptime and the router set the limits. Production for paying clients should run somewhere with an SLA.
- Mail is the one thing I wouldn’t self-host. Deliverability is a full-time job: SPF, DKIM, DMARC and a clean IP reputation all have to be in place before Gmail trusts your mail. That’s why mail went to a provider and not into the cupboard. Know the limits of the cheap plan, though: Migadu’s smallest plan allows 20 outgoing messages a day.
Who it makes sense for
If you’re a freelancer or developer who mostly needs room to build, test and show work to clients, a cheap box at home combined with Vercel, Cloudflare and a mail provider gives you more control for less money than a classic web host.
If you need a site that must be online at 3 a.m. on a Sunday, pay someone to keep it running.
