Coolify

Promptfree, for everyone, and the only version there is

Copy the prompt for Claude CodeCopy the prompt for CodexCopy the prompt for Cursor

Copy

Build me the deployment setup I actually need instead of paying for Coolify's hosted control plane: git push, build, run, renew certificates, on my own servers.

Read this first: the product you are replacing is open source and self-hosting it is free. What the fee buys is a hosted control plane, so the thing that deploys your servers is not itself on one of them — which is a real concern the moment you have more than one. Build this when you have one or two servers and a handful of services, which is where most people are.

STACK
- Bash and systemd on the server, plus a small Node control process if you want a web interface
- Docker or Podman for the services themselves
- Caddy in front for TLS and routing
- SQLite through better-sqlite3 for deploy history, if there is a control process

THE SHAPE OF IT
- A release directory per deploy: /srv/<app>/releases/<timestamp>, with /srv/<app>/current a symlink to the live one
- Shared state — the database, uploads, .env — lives in /srv/<app>/shared and is symlinked into each release. Nothing that must survive a deploy is ever inside a release directory
- Deploy is: fetch, build into a new release, run migrations, switch the symlink, restart, health check. If any step fails, the symlink never moves and the old release keeps serving
- Rollback is moving the symlink back and restarting. It must be one command and it must be tested before it is needed
- Keep the last five releases and delete the rest

THE DEPLOY SCRIPT, WHICH IS THE WHOLE PRODUCT
- Refuses to run on a dirty tree or the wrong branch
- Builds, and if the build fails the running service is untouched. This is the single most important property: a failed deploy must be a non-event
- Runs migrations before the switch, each once, in order, inside a transaction where the database allows it
- Health check after restart against a real endpoint that touches the database, with a timeout and an automatic rollback if it does not come up
- Logs what it did, with a timestamp, to a file that is kept
- One command, no arguments, no decisions to make at three in the morning

TLS AND ROUTING
- Caddy with automatic certificates, which is a two-line configuration and the reason this part is not hard
- One site block per service, reverse proxying to a local port
- On-demand TLS if customer domains are involved, with an ask endpoint that checks the domain against your own database — without that check, anyone pointing a domain at you can make you request certificates on their behalf
- Security headers and a content security policy set once, in Caddy, for everything

SERVICES ALONGSIDE
- Postgres, Redis, a database of some kind — each as a container with a named volume, a memory limit, a restart policy and a health check
- Never expose their ports on the public interface. Bind to localhost or to a container network, and say why: an open database port is found by a scanner within the hour
- Backups of every stateful service to a script that runs nightly, uploads off the machine, and is restored from once as a test

WEBHOOKS AND AUTOMATION
- A small endpoint that verifies a signature from the forge and triggers a deploy for a matching branch
- Deploys serialised per app — two at once is how you get a half-built release
- The result reported back, and a notification on failure
- A manual deploy path that does not depend on the webhook working

MONITORING THE BASICS
- Systemd units with Restart=on-failure and a journal
- Disk, memory and certificate expiry checked by a timer, alerting by email
- An uptime check from somewhere that is not this machine — a free external service is fine, and it is the one thing that must not be self-hosted here
- Log rotation configured before it is needed, not after the disk fills

THE HONEST PART
- Write in the README what this does not do: no cluster, no zero-downtime rolling deploy, no automatic scaling. A restart is a second of downtime, and for almost everything that is acceptable
- Write down what breaks when the server dies, and how long the restore takes. If you have not timed it, you do not know it
- The control plane, if you build one, should be reachable when the app is not — and ideally from a different machine

WHAT MATTERS MOST
The failed deploy and the rollback. Break the build deliberately, deploy, and confirm the site never went down; then roll back and time it. Everything else here is convenience, and this is the part that decides whether you deploy on a Friday.

Give me the deploy script, the rollback script, the systemd units, the Caddy configuration, the backup and restore scripts, and a README that reads as a runbook rather than an introduction.