Practical tools for system, network, security, and operations administrators.

SYS-C0771 · Web Server Guides

Caddy Configuration Guide

Serve a Caddy lab site on loopback without asking the public CA for a name you do not own.

Updated 25 October 2026 · Pillar guide

This is an original planning guide. It is not a vendor manual, an official standard, or legal advice. Example addresses use documentation ranges. Review every command before you paste it.

Where Caddy sits

Caddy reads a Caddyfile. A site address, a root or a reverse_proxy, and a reload after caddy validate. The product will try to obtain a public certificate when a site block looks like a public name. In the lab, do not give it a public name. Use http://127.0.0.1:8080 or the local TLS mode the version documents, so no certificate request leaves the building.

Automatic HTTPS is a production feature for a name you control, after DNS already points at this host and you intend to renew. It is not a side effect of copying a sample that says example.com. example.com is documentation. Pointing a real unused name at Caddy and letting it fetch a certificate is a decision you write down, not an accident.

reverse_proxy points at an application you run. The upstream is in the ticket. The web guide owns the public-listener decision. This page owns the Caddyfile.

The Caddy generators draft a site, a proxy, and a TLS shape. Read the site address first. If it is not loopback in the lab, edit it before you start the service.

A loopback site you can fetch

Step 1 for Caddy is the address. Set the site to http://127.0.0.1:8080 and a file_server root or a static file. A public hostname in the lab triggers a certificate attempt. Check: No certificate request is logged for this start. If the check fails, do not reload again and do not point a name at this host. Restore the previous file. The generators for Caddy draft a site in the browser and do not open a firewall. Delete every listen-on-every-address line for the lab. Replace prompt-for-this. A config from a neighbouring web server is not this syntax.

Step 2 for Caddy is validate. Run caddy validate on the file. A reload of an invalid file is a guess. Check: Validation succeeds. If the check fails, do not reload again and do not point a name at this host. Restore the previous file. The generators for Caddy draft a site in the browser and do not open a firewall. Delete every listen-on-every-address line for the lab. Replace prompt-for-this. A config from a neighbouring web server is not this syntax.

Step 3 for Caddy is the fetch. Start or reload and fetch the test page from the host. The proof is the page, not the process list alone. Check: The page body is the one you wrote. If the check fails, do not reload again and do not point a name at this host. Restore the previous file. The generators for Caddy draft a site in the browser and do not open a firewall. Delete every listen-on-every-address line for the lab. Replace prompt-for-this. A config from a neighbouring web server is not this syntax.

Step 4 for Caddy is proxy later. Add reverse_proxy only for an upstream you already fetched directly. Caddy cannot fix an upstream that is down. Check: The proxied path returns the upstream's test response. If the check fails, do not reload again and do not point a name at this host. Restore the previous file. The generators for Caddy draft a site in the browser and do not open a firewall. Delete every listen-on-every-address line for the lab. Replace prompt-for-this. A config from a neighbouring web server is not this syntax.

Step 5 for Caddy is the draft. Compare the generator draft to the file you loaded. Drafts often listen more widely than the lab. Check: The loaded file matches the ticket. If the check fails, do not reload again and do not point a name at this host. Restore the previous file. The generators for Caddy draft a site in the browser and do not open a firewall. Delete every listen-on-every-address line for the lab. Replace prompt-for-this. A config from a neighbouring web server is not this syntax.

# Caddy lab. Review before you paste.
# Site address: http://127.0.0.1:8080
# Do not use a public name in the lab.
# caddy validate before reload.
# reverse_proxy only to an upstream you run.
# example.com in a sample is documentation, not your site.

Certificates come after the name is yours

Public certificates wait until the name and the DNS are true. Local TLS stays on the lab host. The certificate store Caddy uses is not a place to drop a key you also paste into chat.

Validate before reload. A Caddyfile that fails validation does not get a restart to see what happens.

When you do turn on public certificates, the rollback is the previous Caddyfile that served the lab, and the DNS change is the migration guide's problem.

What not to paste

Caddy is not Nginx. Do not paste a server block.

A sample that names a public site is not a lab file.

Automatic HTTPS is not mandatory for a test page on loopback.

The hour you stop

Write the hostname, the role, and the clock time before you change Caddy. If this host is also serving users, it is not the lab and this first hour belongs on a spare. The provisioning guide is how the spare gets built. A change without a spare is a change you can still postpone.

Keep the console or the out-of-band path open for the whole hour. Reloading a network service, a firewall, or a secure-shell daemon from the only session is how a lab becomes a drive across town. The hardware guide names that path for a physical machine. A virtual machine uses the hypervisor console. A cloud instance uses the serial console the platform gives you. Test it before the reload, not after.

The address 192.0.2.10/24 and the gateway 192.0.2.1 are documentation. They belong in the lab note and nowhere else. On a real network, use the prefix the network desk already assigned. On a cloud instance, use the address the platform assigned and do not hard-code the documentation range into user-data. If a generator filled the documentation address into a production draft, delete it.

prompt-for-this is a reminder that a secret is missing. It is not a password you set, not a token you commit, and not a string you leave in a configuration file because the service started. The real secret lives in the password store. The ticket says who can read it. A file that still contains the reminder is a file you have not finished.

One service is enough for the first hour on Caddy. The purpose from the fundamentals page says which service. Every other package stays uninstalled until its own guide has a ticket. A host that enables a web server, a database, and a queue in the first hour has three changes and no proof. Finish the login, the address, and the update. Then stop.

Read the package transaction before you accept it. A count of packages is not a review. If the list contains a repository you did not add, stop. If it contains a kernel, the patch guide's reboot decision applies and the console stays open. If it contains software this role does not need, remove that line from the plan rather than installing it and disabling it later.

The previous image, the previous configuration file, or the previous package version is the rollback. Write the path down while the host still works. A rollback you invent after a failed reload is a guess. The migration guide assumes the old copy is still there. Do not delete it in the same hour to free a disk.

When the checks pass, copy the notes into the ticket and leave the host alone. The next host waits until a second person can repeat the login and the address from the notes. A fleet built from memory drifts. A fleet built from the file the provisioning guide stored stays boring, which is the point of Caddy in production.

If a step on Caddy fails twice, stop for the day. The failure is the result. Adding a forum command beside a half-applied config produces a host nobody can rebuild. Rebuild from the image instead. The time you lose is smaller than the time you will spend explaining the hybrid.

Hand the finished lab to the guide that matches the role: web, mail, storage, backup, or hardening. Caddy being installed is not the same as the role being ready. The role guide has its own proof. Do not skip it because the package manager exited without an error.

A second person reads the hostname, the address, and the rollback path aloud before you build the next host. If they cannot find those three in the ticket, the notes are not done. Silence is not agreement. The check is the sentence they repeat back.

Store the working notes next to the image version. A month from now Caddy will still boot from that pair, and the chat where you remembered a flag will be gone. The file is the memory. The host is replaceable. That is the whole point of writing this hour down.

Caddy checks before a public listener

Confirm this before you call the change done: Caddy is the process that is listening, confirmed on the host. The lab site binds to 127.0.0.1, not to every address. The config test passed before the reload. A fetch of the test page from the host itself succeeded. The admin or dashboard port is not reachable from a public address. The document root is readable by the worker and not writable by it. No certificate was requested for a name you do not control. The generator draft was edited until it matched this file. The previous config is saved. prompt-for-this was not left in a password field. Logs from the test do not contain a secret. A second person read the listen address. The Caddyfile site address is loopback. No public certificate was requested. caddy validate passed. example.com was not left as a real site name. The Caddy process user can read the site files and cannot write them. And a reload was done only after validate, and the previous Caddyfile is the rollback. If one of those fails, stop.

Tools for this guide

These tools run in the browser on this site. They do not log into your machines.

Guides in this cluster

These supporting pages belong to this pillar.