You described what you wanted, an AI wrote it, and it works. Then you close the laptop and it stops, or you try to send it to a friend and realise there is nothing to send. This is the point where vibe coding stops being vibes.
The good news is that the remaining work is small and mostly mechanical. The bad news is that it is unfamiliar, and the AI that wrote your code is oddly bad at explaining it, because it is not a coding problem.
First, work out what you actually built
Every deployment guide starts by assuming you know this, which is not fair to someone whose first line of code was a prompt. There are three shapes and they go to different places.
| IF YOUR PROJECT… | IT IS | IT NEEDS |
|---|---|---|
| Is only HTML, CSS and JavaScript in a browser | Static | Any static host, free |
| Has a server that answers requests | Full stack | A host that runs your backend |
| Keeps running with nobody using it | Always-on | A process that never sleeps |
The third one is the one that surprises people. A Discord app, a Telegram app, a scraper on a loop, a thing that watches a price and messages you. None of those are websites. Nobody visits them. They sit there waiting, which means they have to be running while nothing is happening.
The three things that always break
AI-generated projects fail on deploy in the same three ways, near enough every time. None of them mean the code is bad.
- 01Secrets that only exist on your machineYour API key or bot token lives in a .env file that is deliberately not uploaded. The deployed copy has no idea what it is. You set it again on the host. See the environment variables article.
- 02Addresses that point at your own computerAnything saying localhost or 127.0.0.1 means "this machine". On a server, that is the server talking to itself. These need to become real addresses or environment variables.
- 03Things that were installed and never written downIf you installed something to make it work locally, it has to be in package.json or requirements.txt, or the host will not install it. Ask your assistant to check that every import appears there.
That is genuinely most of it. Ask whichever AI wrote the project to go through those three, and it will usually find and fix all of them in one pass, because now it knows what to look for.
Why the free option stopped working
Free tiers are real and often fine. The catch is that most of them sleep: no traffic for fifteen minutes and your project is suspended until someone wakes it. For a website that is a slow first load. For anything always-on it is fatal, because "no traffic" is its normal state.
This is why people report that their app "works for a bit then dies" on a free host. Nothing died. It was put to sleep, on purpose, exactly as documented.
What it should cost
A static site: nothing, permanently. A small always-on project: three to seven dollars a month, and be suspicious of anything claiming otherwise, because a process that runs continuously costs the host money continuously.
You do not need a big machine. Most AI-built projects are small. An app with a few commands sits under 200MB, which is the cheapest tier everywhere.
You do not have to learn a dashboard
The traditional path is: make an account, read the docs, learn their panel, write a config file, push to git, watch a build log you do not understand. That is a lot of new concepts at the exact moment you wanted to be finished.
If your project is always-on, Spocket is the version where you skip that: connect it to Claude, Cursor or whichever editor you built in, and say "deploy this". The assistant sends the folder, sets the variables and reads the logs back to you when something goes wrong. Same loop you have been in all along.
And if you built a website rather than an app, use Vercel or Netlify and ignore us. That is what they are good at, and it is free.