Hacker News new | past | comments | ask | show | jobs | submit login
Show HN: Claude Artifacts but creating real web apps (gptengineer.app)
189 points by antonoo 5 months ago | hide | past | favorite | 48 comments
Hey Hacker News!

Launching gptengineer.app into beta today.

It's like Claude Artifacts, but:

- you can edit the code in your fav IDE (two-way github sync)

- installs npm packages

- automatically picks up build and runtime errors and fixes them

- very fast, built with rust

The full stack capabilities are built on supabase (prefer to not have to handle auth + user data at this point so this is owned by the user)

The seed for this project was an open source experiment, posted about that previously here:

https://news.ycombinator.com/item?id=36422730

Would love feedback if you give it a try!




Wow. I typed in: Create a page that has a snake game clone that you can control with the keyboard.

Within <20 seconds, i was playing it. That was crazy. I then asked it to add better graphics, retro-style. And add a counter displaying the current snake length. And then finally, add a high score tracker after the game ended. Adding the tracker caused an error, and I told it to fix the error. It looked at the logs and fixed it. There is still a bug where the top 10 scores all update at once, but I'm sure if i didn't hit the free credit limit on the site I could ask it to fix that.

I did all of this within ~3 minutes. It works, and while it is very suboptimal (it plays Snake using basic React state code, not ideal lol), the speed of this and how surprised I am reminds me of the first time I talked with ChatGPT 2 years ago. Amazing work.

I don't know if these are shareable, but this is the link I'm using to view the page it built:

https://preview--arrow-snake-adventure.gptengineer.run/


Just a warning about testing with snake/tetris - they're the most reimplemented games in the world and there's so many examples those models learned from. For a more realistic test, it's better to try something novel.


Mm, even so Tetris is one of my tests, and one of the models small enough to run locally on my machine… started off badly, then suddenly switched from "Tetris as a web app" to "train an ML model in python".

So I think it's a valid quick sanity test, like fizz-buzz for human devs.


You can get government subsidies in Germany using snake :)

News in German from a week ago, 180,000 Euros in Government subsidies to develop a modern Snake: https://www.iphone-ticker.de/spielefoerderung-in-der-kritik-...


Agreed, but I think that dismisses the ease of use of this product compared to other offerings. Being able to type in a single sentence and have a full webpage of exactly what you asked for in less time than takes to stand up and stretch is kind of magical.


search engine dot com and “snake game” is still probably faster though. which is the point gp was trying to make. there’s so much training data on snake, that this is what you’re essentially doing.


may be search it's faster, but this thing is supposed to do it something like fortran and console. ;)


Just want to point out that the flow you're describing is not that noteworthy and has been doable in many instant text-to-app services for at least a year, and doesn't really include any of the points that OP highlights as the USP. Vercel has their own (v0) etc. Your version even looks the same as when you tell WebSim to make a snake game. Townie from val.town is the best in my opinion because it can also generate live backend endpoints.

It's a really saturated market but I wish OP the best of luck.


This feels like such an inevitable product. I already love using it. Congrats on the excellent launch!

One feature request would be a search for the existing public projects so that I can use remix to jump start my project.

Two questions:

1) Given the name, is it fair to assume that this uses OpenAI's API as opposed to Anthropic or others?

2) In the initial create prompt, what kind of files are allowed as attachments? What type of files give best results?


1) We use whichever LLM that is performing best at the time, so no, not necessarily OpenAI 2) Only images as attachments for now. You can add more files by uploading them to the connected Github repo.


Well done. Very nice.

I really like the UI/UX, and admit that I am a bit jealous (as langcss.com creator) of how polished it is. I like the way it tells you what files it is working on but doesn't stick a full IDE in your face.

I am not worried this year about AI taking my job - it couldn't manage to capture the camera for my example app. It did get the permissons. Still this isn't useless - it can be easier to debug a non-working app than build from scratch.

So at this point you need a coder! But that will change pretty quickly over the next 2-5 years I am sure.

There is certainly a danger that this kind of app could do a lot of jobs, especially once mistakes are limited so you don't need an Engineer to come fix it. However learning some non technical skills like business analysis is probably wise for most of us :-).


> especially once mistakes are limited so you don't need an Engineer to come fix it

The main issue here is "are there mistakes, in the first place?".

Maybe people will become tolerant to machine mistakes. Engineers make mistakes, but it's expected. And the kinds of mistakes engineers make are human.

Machines make mistakes that look silly. The kind of mistake "an engineer wouldn't make".

So these apps will probably tackle an unserved demand (things not valuable to justify an engineer). Anything else I predict will continue to use engineers - more productive with AI support.


Thanks for the kind words!

Agree on the timelines here. This is currently a tool.


This has the same problem as the vercel v0 demo. This is good for like basic apps and stuff but nothing major beyond that and then if you want toy apps in the first place, to cook it up, if you know little bit of HTML and CSS, is not that hard either.


Well, yeah, that’s what it’s for: something for PMs/POs to make sketches with - in even less time than firing up Figma or Xd (…do people still use Xd?)


Everyone super excited. :) In two years everyone has been laid off. ;)


Everyone was going to be laid off because of AI within two years in 2022. Well, many have been, but technically not because of AI. Well, not because AI took over their jobs proper, anyways.

AI used to always be 5-10 years away. Now AI is here. What's 5-10 years away now is all the reconfigured or reconstructed economic infrastructure around it.

It may be easy to convince me that tools like this can take over development of web apps. It's harder to convince me that, in the timeframe where these tools actually eliminate web app jobs, that we will be worried about our web app building capacity.

The most generalized economic pipeline is (labor + capital) -> production -> welfare. We're obviously going to be reconfiguring a lot of labor, a lot of capital is going buh-bye, much of production will look nothing like it used to, and we haven't even started discussing whose welfare we're going to be maximizing (hopefully it's people in general rather than real estate owners.)


That's similar to my thinking. It seems we're still (or again?) in an age where an incredible amount of software is being written to validate value propositions and business models. Does anybody need something like this? Does anybody want something like this? What would people pay for it? Surprisingly often, this isn't done too well, and a lot of devs are thrown into working on stuff that is unlikely to be useful to anyone mid term. Many of us are just building lottery tickets.

Tools like this can potentially lead to this on absolute steroids (even as less capital is available), insane amounts of experiments around new products, features and businesses. That way we'd discover more valuable software, and need more developers to scale and maintain that stuff. Because the economics do reverse: At some point in a product's life cycle, that dev salary is well within the profit margin, and not really worth saving. Quite different from the experimentation phase where quantity>quality.

Another thing that could happen is that individuals build their own software more, like Excel on steroids. There would still be value in solving problems for users at scale, but this whole experimentation process wouldn't be so breadth-first anymore. And I suppose funding would be pretty dry for companies "just" solving one problem.

"AI" isn't even the biggest factor here, I believe. Tools to quickly run experiments and cheap labour are readily available. It kinda seems the times where every company felt they need to hire as many developers as possible are pretty much over.

In both scenarios, I suppose good times might be ahead for competent developers that can earn a client's/CEO's trust, and/or excel at solving problems automated tools fail at. Bad times might be ahead for anyone who just executes experiments other people came up with. Unfortunately, I fear that's the majority of us. It might just get more painful before our industry normalises after this prolonged phase of non-stop growth. But I do believe it will. Demographic change leading to shrinking work forces in most developed countries should help soften the blow.


We systematically get posts and blogs and real-world experience showing that software development is actually about maintenance.

Yesterday we had an article about so many half-arsed software in maintenance mode that the world would be ending soon, crumbling under the weight of not-maintained enough software.

So far I've seen LLMs (I pay for GTP 4o btw) very good at producing boilerplate code. Does it help? Heck, yup.

Does it produce even more code needing ever more maintenance? You bet so.

At this rate it looks like we're going to need way more software developers, not less.


Two years is good for me. So long as my tenant isn't also laid off and the stock market doesn't crash as a result of this… or at least everyone gets UBI.


This is mind blowing. This is not just a toy, I made a landing page that is publishable in 3 prompts.


Thanks! Happy to get more things not working as expected yet if there's anything.


Wow really nice. I am wondering how easy is it then to transfer the code to a webhoster and manage packages, update and security risks?

I am not a developer (can edit a few lines) but would be a target customer for these. I just wonder how to manage these risks compared to let’s say a CMS.


So it's already hosted for you, you don't have to do anything to get a production ready URL.

But yeah if you want your own hosting for some reason you own the code and can take it to a provider like Netlify or Github Pages fairly easily.

Send me an email at viktor@lovable.dev if you have any questions :)


Gave it a whirl and it certainly did the job, albeit a bit basic in its implementation. Would love to see it veer into being a bit more creative in its choices (maybe you can have an input for a more chaotic input like they do in midjourney. Congrats on launch! This is very cool


Yeah the standard shadcn setup tends to be a bit boring. But if you ask it to be a bit more creative you certainly can get it to make some pretty fun stuff: - https://preview--retro-hacker-portal.gptengineer.run/magnus (source: https://gptengineer.app/projects/91b06c8d-16a8-4892-8cb5-d91...)


Is this like https://websim.ai/ but more explicit in terms of what it's doing and it's purpose? I have used websim.ai for fun and there are quite a few interesting websites there. But what they are doing is all disguised under their address bar.


what does disguised under address bar mean?


all they give you is an address bar where you are supposed to write address of an imaginary website and it will bring that website into existence for you. It does a little more too, instead of url, you can write prompt in the address bar, and once it makes a website for you, again in the address bar you can prompt it to modify the website to your will.


I see a comment saying it used "basic React state code". So its output is based on React as well as Supabase? Or might it uses different base technologies depending on what it thinks is best for what it's building?


This is great, now do mobile!


It's such an easy way to create prototypes for anything you could think of


At the point I run out of free credits, I was still arguing with the AI about stuff it was doing wrong.

I think the amount of free usage needs to be high enough for people to get a small taste of success at least!


Well done, works well indeed. You are still active on the OSS project I see; the new site has been built in Rust, so I guess they diverged quite a bit?


Where do you see this compared with Cursor or Copilot?


Awesome work! How are you installing npm packages & compiling in the browser IDE without any local filesystem?


Maybe a docker spinup for each app (which would be kind of nuts and could get out of hand perhaps) ? I have to admit I didn’t look at the site much so have no idea really


We use a dedicated firecracker VM for each project with a rust binary wrapping a vite dev server.


What does the rust binary do here?


Could be WASM?


Very cool, love the community projects showcase and the ability to remix them.


Thanks!

Let me know if you have any input on what we could add, we're iterating super fast on the exactly right form factor.


Having not used Claude artifacts, maybe a better introduction to how the tool works/looks before signing up?


Thanks, noted


How do I get access?


So we opened up our waitlist so it's open right now: https://gptengineer.app

(Might have to readd the waitlist though if we get too much traffic.)


Having fun so far but given the advice to ask for small incremental changes the 10 runs per day is not enough. Even the Pro level at 100/day for $19/month is too low if I really want to play with it.


love it. works great out of the box




Consider applying for YC's Spring batch! Applications are open till Feb 11.

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: