A backend for vibe coding
You described an app, the AI built it, and it works — on your screen. Vibe coding is real; the demo in your browser tab is not a trick. The trouble starts when you send the link to someone else, because what got generated is a frontend, and a frontend alone can't serve a second person.
What breaks first
The failures arrive in a predictable order:
- The data isn't real. It lives in the browser tab. Refresh and it resets; your colleague opens the app and sees none of what you entered — there's no database behind it.
- Everyone is the same person. No accounts means no "my orders" vs "your orders" — and an app that shows every user everything is one you can't actually give to customers. That's auth plus row-level rules on the data itself.
- Uploads go nowhere. The photo picker works; the photo has no storage to land in.
- There's no address.
localhost:3000isn't a link you can put on a business card. The app needs hosting and a domain, with TLS handled. - Nothing happens on its own. The receipt email, the password reset, the nightly report — those need email and scheduled jobs running server-side, not in a browser tab someone has to leave open.
None of this means the AI did a bad job. It means "build me an app" was quietly two requests: make the thing, and stand up everything behind it.
The second request is what a Wok is
A Wok is the whole live backend in one piece: a Postgres database, auth, storage, realtime, edge functions, a scheduler, email, SMS, hosting on your domain, a model gateway, and key management — provisioned together, private to your team. You don't pick the parts off a menu or wire them to each other; a Wok arrives already wired. The stack goes through each piece.
Closing the gap without changing how you work
If you're building inside Cursor or Claude Code, point it at the TellWang MCP server and keep going — the same session that wrote your frontend can provision the Wok, create the tables, set the access rules, publish the site, and check the live result. There's a step-by-step in Deploy a Claude Code project.
And if the AI chat was the whole point — you never wanted to see code — skip the tools entirely. Tell Wang what you made and what it should do for other people, and Wang builds the backend, connects your app to it, and hands back a URL that works for someone who isn't you. Before it says "done," it opens that URL in a real browser and reads the console and network errors a simple "page loaded" check would miss — the same gap this whole page is about, caught before your users catch it. How Wang works shows the build loop.
Common questions
Why does my AI-generated app break when someone else uses it?
Because it was generated as a frontend. The data lives in the browser tab, there are no accounts so everyone would see the same thing, uploads have nowhere to go, and the only copy runs on your machine. None of that shows up while you're the only user — it all shows up the moment a second person opens it.
Do I need to learn SQL or server administration to fix it?
No. Describe what the app should do and Wang builds the backend it needs — the database, the logins, the storage, the live URL — and tests the result in a real browser before handing it over. If you do write code, your tools can drive the same backend directly over MCP.
Can I keep building in Cursor or Claude Code?
Yes. Point Cursor or Claude Code at the TellWang MCP server (https://tellwang.com/v1/mcp) and the same session that writes your frontend can provision the Wok, create tables, deploy functions, publish the site, and verify it live.
What exactly do I get with a Wok?
One live stack with everything a working app needs: a Postgres database, auth, storage, realtime, edge functions, a scheduler, email, SMS, hosting on your domain, a model gateway, and key management — provisioned together and isolated from every other team's.