*Here’s what it does:*

You load the page. You write 1 prompt and you get a mini app back in under 2 minutes. There’s no sign up, and you can see what everyone’s creating in real-time!

Each mini app comes with it’s own database and backend, so you can build shareable apps that save data.

*What’s different*

There are a lot of app builders that promise you’ll build production software for others. But we think true production software can take a long time to get right. Even if you don’t need to program there’s a lot of work involved.

What if we turned the promise around? Instead of “you vibe code software companies”, it’s “you build fun software for yourself”.

If you cut the problem right, LLMs as they are today can already deliver personal software. manyminiapps is meant to be an experiment to demonstrate this.

You may wonder: do you really need personal software? We’re not 100% sure, but it’s definitely an interesting question. Using manyminiapps so far has been surprising! We thought our friends would just try to build the common todo app, but instead we found them building wedding planners, chord progression helpers, inspiration lists, and retro games.

*How it works*

Instead of spinning up VMs or separate instances per app, we built a multi-tenant graph database on top of 1 large Postgres instance.

All databases live under 1 table, on an EAV table (entity, attribute, value). This makes it so creating an “app” is as light as creating a new row.

If you have heard of EAV tables before, you may know that most Postgres experts will tell you not to use them. Postgres needs to know statistics in order to make efficient query plans. But when you use EAV tables, Postgres can no longer get good statistics. This is usually a bad idea.

But we thought it was worth solving to get a multi-tenant relational database. To solve this problem we started saving our own statistics in a custom table. We use count-min sketches to keep stats about each app’s columns. When a user writes a query, we figure out the indexes to use and get pg_hint_plan to tell Postgres what to do.

*What we’ve learned so far*

We’ve tried both GPT 5, Claude Opus, and Claude Sonnet for LLM providers.

GPT 5 followed the instructions the best amongst the models. Even if you told it a completely nonsensical prompt (like “absda”, it would follow the system prompt and make an app for you. But GPT 5 was also the “most lazy”. The apps that came out tended to feel too simple, with little UI detail.

Both Claude Opus and Sonnet were less good at following instructions. Even when we told them to return just the code, they wanted to returned markdown blocks. But, after parsing through those blocks, the resulting apps felt much better.

To our surprise, we didn’t notice a difference in quality from Opus and Sonnet. Both models did well, with perhaps Sonnet following instructions more closely.

To get good results we iterated on prompts. We initially tried giving point-by-point instructions, but found that a prompt with a full example tended to do better. Here’s what we landed on:

https://gist.github.com/stopachka/a6b07e1e6daeb85fa7c9555d8f...

Let us know what you think, and hope you have fun : )