Moving 35 websites to a new server in a weekend, with Claude Code

Thirty-five sites, from a 2012 CMS on PHP 5.6 to Drupal 11, moved off a 17-year-old managed server. Every page was checked on the new server before its address changed.

For seventeen years every website I host lived on one managed server. This weekend all 35 of them moved to a new one: a 2012 CMS that needs PHP 5.6, two Drupal 11 sites, WordPress, four Club 100 apps, fifteen client review copies and this site. The first went live on the new machine at half past one on Saturday afternoon and the last at half past nine on Sunday morning. The monthly bill fell by about 40%. Claude Code did most of the typing. This is how the move was planned and checked, and where the risks were.

Illustration: small website cards flowing along an arc from an old cluttered server tower to a clean modern server
Thirty-five sites, one weekend, every page checked on the new server before its address changed.

Why move at all

The old machine was a managed dedicated server, rented since 2009 and very reliable. Managed meant the host looked after the operating system and I never had root access. That suited the sites of 2009 and had become a ceiling by 2026:

  • No root access. Nothing could be installed beyond what the host offered, so any search service, queue worker or newer database engine needed a second server.
  • A file-count quota. The hosting panel capped the account at 1,048,576 files. A Drupal 11 install is 40,000 to 45,000 files, so a few review copies of client upgrades pushed the account to its limit and old copies had to be deleted first.
  • One user for everything. Every site ran as the same account, so a flaw in one old application could reach the files of all the others.
  • Price. £90.74 a month for a server that measured at about 1.5 GB of its 16 GB of memory in use.

The replacement is a self-managed VPS with the same host, matching the old server's 8 cores and 16 GB of memory, with 480 GB of NVMe storage and root access, at £41 a month. Mailboxes used to come with the managed server and are now a separate pack, which brings the total to about £54 a month. That is roughly £440 a year saved, on a machine that can do more.

One site, one Linux user

The first afternoon went on the base server, written as scripts so that it can be rebuilt the same way: Ubuntu 24.04 LTS, Apache with PHP-FPM, MariaDB, a firewall that only opens SSH and the web ports, fail2ban, automatic security updates with a 4 am reboot when the kernel needs one, and SSH keys only, with no passwords and no root login.

The most important decision was isolation. One command, vpsg-site-add, creates a site with its own Linux user, its own PHP pool running as that user, its own database and its own Apache virtual host. A site's PHP cannot read another site's files. That matters a lot on a server that has to run a 2012 application next to a 2026 one, because the old one cannot drag the others down with it.

PHP 5.6, 7.4, 8.3 and 8.4 are installed side by side, and each site chooses its version. The oldest site still runs on 5.6 for now. The others moved up to the newest version their code supports.

The rule for every site: copy, prove, switch

Thirty-five sites gave thirty-five chances to break something in public. Each one went through the same sequence, and no site's address was changed until it had passed:

  1. Copy the files and database straight from the old server to the new one. Server to server is fast: a 3 GB site took 33 seconds.
  2. Set the site up on the new server with its PHP version, its database connection pointed at the local copy and any settings the old host had applied silently.
  3. Prove it. Before any DNS change, every page of the site was fetched from the new server while the public address still pointed at the old one, and the two answers were compared. For this site that was 106 of 106 pages matching. HTTPS was tested too, with a temporary certificate, so a certificate problem could not appear after the switch.
  4. Switch the address. The certificate for the real name was then issued in seconds.
  5. Retire the old copy visibly. Once the new copy was confirmed, the old one was set to answer "this site has moved". Nothing could keep using it without anyone noticing.

Why the old copy fails on purpose. After a move, the dangerous case is a site that appears to work but is still talking to the old server: a cached address, a hard-coded IP, a cron job nobody remembers. Making the old copy refuse every request turns that silent case into an obvious one, while both servers are still there to fix it.

Six things that would have broken quietly

None of these show up on a home page check. Each one would have looked fine on the day and caused a problem weeks later.

1. Mail from the website

A managed server has a mail system built in, so PHP's mail() function just works. A new VPS has none. Every contact form, password reset and order email that relies on mail() fails silently: the visitor sees "thank you", and the message goes nowhere. Each site that sends mail now sends it through a real mailbox over authenticated SMTP, set up in that site's own PHP pool.

2. The IPv6 half of the address

Most of the domains had an IPv6 record as well as the usual IPv4 one, both pointing at the old server. The new server is IPv4 only. Change the IPv4 record and forget the IPv6 one, and visitors on networks that prefer IPv6 keep reaching the old server, which is about to refuse them. Both records changed together on every domain.

3. Enquiries arriving during the move

Some sites store form submissions in their database. A copy taken in the morning and switched over in the evening loses whatever arrived in between. Those sites had a fresh database copy taken immediately before the switch, and the submission count on the old server was checked again later to confirm nothing had arrived after the copy.

4. Settings the old host applied quietly

Old PHP code sometimes depends on settings that were on by default in 2012, such as short <? tags. On the old server that was simply how things behaved. On a clean server those pages show source code. Settings like that are set per site, in that site's own pool, so they never apply to the modern sites.

5. Ten names, one certificate

This site answers to ten names: three domains with and without www, plus older domains that point here. Requesting a certificate per name gives a virtual host that only presents one of them. The fix is one certificate covering all ten names, issued in one request.

6. The old host's limits during the move

The old server started refusing connections when the new one opened too many in quick succession. The answer was to put each site's copy into a single connection and to leave a gap between batches.

What Claude Code did, and what I did

Claude Code ran in a terminal on my workstation with SSH access to both servers through its own key and user on the new one. It wrote the setup scripts, ran the copies, fixed each site's configuration, carried out the page-by-page comparisons, and changed the DNS records in the host's control panel through the Claude in Chrome browser extension, in a browser session I had logged into myself. It kept a running log of every site, so the state of the move could be picked up from any point.

The decisions stayed with me: which sites to keep and which to retire, when each one switched, which mailbox a site sends from, and anything to do with passwords. When something odd turned up on a site, Claude Code stopped and asked rather than tidying it away. A few old folders turned out to be in use by someone, and a few that looked important were not.

It was a long day and a short night of supervision, not a hands-off one. Doing the same move by hand, at the same standard of checking, would have taken me a working week. The checking is the part that would have been cut.

After the move

A server you manage yourself needs the things the managed host did quietly. These went in before the last site moved:

  • Health checks every five minutes, with a phone alert when a site, the database or the disk has a problem, and another when it recovers.
  • Nightly database dumps kept for 14 days on the server, and a daily copy of files and dumps pulled to a disk in my office.
  • A browser check every three hours that loads each site's home page in a real headless browser and raises an alert if a title changes, a password prompt disappears or a page errors.
  • Deployment and security-update tooling repointed at the new server, so a Drupal security release still goes out the same day.

Google PageSpeed Insights on this site the morning after: 100 for accessibility, best practices and SEO on every page tested, and 98 to 100 for performance on mobile.

If your sites are on a server like that

Plenty of agencies and small businesses have a server just like the old one: rented years ago, still running, holding a mix of sites of different ages, and costing more each year. Moving it is a well-bounded job if every site is proved before its address changes. I can do that for your sites, with the same checks and a fixed price agreed first. Agencies can have it done for a client under Drupal for agencies, and old Drupal sites can be moved and upgraded together through Drupal 7 support. For anything else, get in touch.

© 2026 Graith Internet.

Graith Internet is a UK web development company specialising in PHP, Drupal, and AI-assisted development.

Find this useful? Tell Google to show you more from Graith Internet in Search, Discover and News.