*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 : )
The result is kind of a nice quality too: https://vital-star-5z6w40.manyminiapps.com/
Great work! Any future plans on this?
Some nice examples:
A collaborative pixel-based image builder: https://peace-robin-a2dw95.manyminiapps.com/
Analog synthesizer controlled by accelerometer: https://magic-gold-z13d9k.manyminiapps.com/
Here’s a bear that poops prime numbers.
https://light-whale-5opfal.manyminiapps.com
Well done guys.
https://hope-rock-071r26.manyminiapps.com
It decided which celebrities to use and also sourced photos for some of them
The everything in one EAV table concept sounds very interesting as well. Can you say more about that? What kind of queries does it support? Sounds a bit like datomic or datascript. How do you tell the agent to use it?
Yes, we definitely take inspiration from datomic and datascript, as well as Tao (Facebook's DB). EAVs have a few good properties. First, it's more natural to do last-write-wins at the column level this way. Second, we get a real shot at creating a multi-tenant database.
With EAVs you can either query them with datalog, or use SQL CTEs. Fun fact, on Instant's backend, we use a datalog-like langauge as intermediary representation, which ends up converting to a SQL CTE.
Technically, you can make query on these as you can on a traditional SQL DB.
To interact with all this, users (and in proxy agents) use a graphql-like language, called InstaQL. It's like if GraphQL was represented as plain javascript objects.
Here's the system prompt, which shows these off:
https://gist.github.com/stopachka/a6b07e1e6daeb85fa7c9555d8f...
It's about 260 lines of home-grown code. Here's a snippet of how it looks:
https://gist.github.com/stopachka/a38f00545d048661cf15a3cf4d...
Two interesting things there:
1. Since Instant is reactive, we use it to convey the stream to the client. This way all chats are "resumable" by default, even if the user refreshes the page.
2. To make sure Instant writes don't block the stream, we use a "natural" buffer approach. Basically tokens from the LLM buffer up during every save.
Alas, we had to add sign in, and scope sign ins to users with an @gmail.com account.
We got a few trolls who were really motivated to use up as many tokens as possible.
We tried blocking Eastern Europe and Indonesia, but that wasn't enough. Hopefully the sign up will help alleviate this a bit.
Cool tool however. Very coherent output given how I find AI.
One question however, if op or anyone knows, any good way/at all to feed a UI and UI interactions back into a model?
Lesson learned, we just added a content moderation filter!
> any good way/at all to feed a UI and UI interactions back into a model?
Right now no, but it's definitely something we'd want to explore. That, and "forking" could be interesting.
Each mini app runs on a subdomain, and the code is evaluated on the local machine. Since a few hackers are interested in the LLM details: Right now we are running Opus 4.1, but we might switch it.
an application to create a database and API of Sylvanian families and track which ones you own and want: https://lush-swan-w7nqsq.manyminiapps.com/
interesting to see that the first app I created allowed user submissions and within seconds there was a slur submitted but now that submission has been removed!
Uncaught QueryValidationError: At path 'worlds.snapshots.$': 'limit' can only be used on top-level namespaces. It cannot be used in nested queries. NextJS 7 tP tL tq tq tV subscribeQuery subscribeQuery 122-677938205b479825.js:1:45042 NextJS 128 tP tL tq tq tV subscribeQuery subscribeQuery a av o0 ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS ux uS
I updated the system prompt a bit, here's a version of your app that's working with it:
https://brave-elm-7m6ar8.manyminiapps.com/
(The chat: https://www.manyminiapps.com/?c=1a5d3481-d2e2-414c-a7c3-4d0c...)
---
For the curious, the issue was that the LLM originally decided to write the query:
```
const { data } = db.useQuery({
worlds: {
$: { order: { createdAt: 'desc' } },
snapshots: {
$: { order: { timestamp: 'desc' }, limit: 1 }
}
}
});
But we don't support limits on subqueries yet.
We had a specific note in the system prompt warning the LLM to avoid this. Interestingly, _removing_ the warning fixed the issue. I guess it's kind of like telling someone to not think of a red elephant.
it took a bunch of tries to get it to do what I wanted, but https://neat-otter-bj288v.manyminiapps.com a crochet spiral counter helper https://calm-puma-bqqdmr.manyminiapps.com - a falling sand game
I'm quite impressed
Instead of the movie poster, it decided on its own to use icons that represent the movie instead (a pill represents Matrix, a gun Goodfellas etc.). I was surprised by its creativity to get around any copyright issues.
The initial movie needs to be dragged on the thin white line (not exactly good ux). Perhaps it's not visible?
Also, if you have dragged the first movie and place the second movie incorrectly, it will be placed back in the "stack"
> build a bingo board app for a group of friends. the owner enters texts (options) in a textbox.
> A link is generated that can be sent to participants.
> They enter their name and can can mark bingo options (5x5 with joker at center) and see a scoreboard.It worked really well, and it was very fast, especially considering that it is on the front page of HN. Congrats, very cool project
Link to the app https://vital-crow-mmrs4s.manyminiapps.com/
I was curious how GPT-5 would do with your prompt. Here's what it got:
http://noble-orca-yq331e.manyminiapps.com/
It got something, but _much_ less creative then Claude in this instance.
https://www.manyminiapps.com/?c=0a043d67-9737-4295-98d5-07d6...
>Build a simple lofi black and white multiplayer doom clone where players drop in near each other in a procedural generated forest world. The graphics should be 2d in a 3d world. Players enter their name before joining. The DB tracks players names, scores, deaths, total play time. When a players is shot too many times and dies they randomly respawn. Players should spawn within visual range of each other. add some harmless non-players animals that roam around. generate some forest background noise to play. players should make a little sound when they walk that other players can hear. players should be able to hide behind trees. Try to make it fun.
It basically worked. https://www.manyminiapps.com/?c=da20213e-6832-4cd8-ac73-7669...
Albeit slowly.
tried to recreate itself
Ah, wish there was a way it had access to babel/runtime. [1]
[1] Though then the mini app builder inside the mini app builder would have a security vulernability, if the mini mini apps actually evaluated.
Played it a bit and found a nice combo:
Heading: JetBrains Mono 400 @ 48px Body: Merriweather 300 @ 16px
Well done.
Then what?
Technically you could use the apps for yourself too, though everything is public so be wary.
Step 2. Scale unicorn SaaS to $1B valuation
Step 3. Sell it, become rich, buy superyacht