In order to keep track of the entire AI space, I find myself jumping around to experiment with different services. I’ve been using GitHub CoPilot ever since it launched, and the industry is rapidly evolving. Now we have Cursor and Windsurf, which are VSCode forks with coding AI agency implemented.
On the rise are fully automated coding Agents directly in the browser like Bolt.new, Replit, and the new Swedish player, Lovable.
I figured, let’s see what all the fuss is about. I had a new idea for a project that I thought would be a good opportunity to test this on. At least for the MVP (Minimal Viable Product) while the scale is small enough.
I tried it and there’s some good and some not so good thoughts I have on it.
What is Lovable?
In short, Lovable is an AI agent that can code entire projects just by having the user tell it what they want. It basically starts with a chat window, just like ChatGPT. The user writes the prompt and Lovable then creates a web application based on the prompt. This approach or concept is what people now are referring to as “Vibe Coding“.
You can see what people have build with it here.
Lovable started as “gpt-engineer” but has been rebranded to Lovable.
Here are some quotes directly from their website:
In mid 2023, before code generation was mainstream, we created gpt-engineer – an open source project exploring how AI could help build software. It became the fastest growing code repository on GitHub to date, and today has over 50,000 stars.
After gpt-engineer blew up, we wanted to make it more accessible, since the open source tool is meant for developers using their terminal. Therefore we built a commercial web version called gptengineer.app, meant to be used by non-technical users. Around the same time in late 2023, we founded Lovable to tackle the larger mission of creating the last piece of software (software that creates other software).
The initial setup
It is really cool to see that with a well enough thought out prompt, you can really get a tailor made boilerplate project up and running. This part impressed me quite a bit.
Just for transparency, I did some research on how to get the most out of a tool like Lovable. How you should structure your initial prompt, what to include, etc.
The prompt I wrote contained the general idea, some of the initial requirements from a UX perspective, also some technical design requirements. Now, Lovable in particular works with ReactJS combined with Supabase backend. It’s a great combination for a MVP so I’m completely ok with this. Don’t think it would be fair to ask for more, eg. like a more advanced NextJS application with a back-end of my choice.
Once I got my initial project up and running, it was time to dive into continued development of the project. Here is where things can fall apart quickly if you aren’t careful. And here is were I’m starting to get a bit more cynical of these “Vibe Coding” services.
The developer experience
When you have your code base done enough for you to start running the project and continue development, you start to prompt for changes, new features, tweaks, etc. Sometimes Lovable got it right. Other times it completely changed the code base to something that broke the entire project.
Sometimes, or quite often actually, Lovable got stuck in a loop where it just couldn’t fix the problem. It wasn’t even close to a solution. Yet it would always confidently claim that it had fixed the issue, even when it clearly hadn’t. Imagine not being a developer while stuck in this loop, it would be a nightmare and cost you a lot of money trying to prompt yourself out of it (as you have limited credit use based on amount of prompts, and after a certain amount of usage, you have to pay more).
Here is where I had to step in and help Lovable to debug things. Fortunately, I’m also an experienced developer and do this for a living. I’m also educated in how LLM’s work. So this was actually expected.
Limitations
This is where I enter my second gripe with web based services like Lovable. The ability to manually continue work does have a lot of friction. The way Lovable works is that you connect Lovable to your GitHub repository and Lovable directly commits and pushes the changes to your branch. So when I needed to do something manually from Lovable I had to check the code on GitHub. In those cases, I just had a fresh version of the code on my local machine and used my favorite IDE to do the manual work. Issue here is the disconnect between the Lovable service and my hands on approach on my local machine.
However, good news for Lovable users. They have recently implemented a web based code editor directly in the Lovable interface. Which is a great step towards solving this issue. But it’s no where near the same developer experience I’m used to. And it’s based in the browser, which is not ideal.
Another annoying thing developing with Lovable (this is actually an issue with LLM’s in general) is the context window. The memory of the AI. Sometimes I worked on bigger features that required multiple steps. Even when I tried to break things down into small steps and always included context at each stage, the AI would suddenly forget the end goal. Instead, it would veer off into a completely random direction.
When it does this, it also feels a bit random. I can’t confirm this is the case, but I suspect that either Lovable or the AI provider is tweaking the context window (how much they allow the AI to remember). So I always need to be mindful of this.
Setting up different development environments can also be a challenge as some of the configuration files are automatically generated. And being a SPA ReactJS application (Single-page application, rendered on the client side) also means reading environment variables can be a challenge.
Who would I recommend this for?
For experienced developers, why not jump in and test it. Just don’t expect it to be a magical thing that doesn’t need your involvement.
For non coders, great to start with something small. But if you eg. are a solo founder with no technical skill. Great for a mockup site with limited functionality, but you should really get a developer involved for further development of your product if you don’t have the skills yourself.
This really won’t replace a developer. Especially if your product idea has any kind of technical complexity to it. And if it doesn’t, then you have to ask yourself, is your idea even worth pursuing?
Final Thoughts
I’m impressed by how quickly Lovable can transform a well-crafted prompt into a functioning MVP. The speed at which you can go from concept to working code is genuinely remarkable, and I see why many developers and founders are excited about tools like this. Among the browser-based AI coding agents I’ve tested, Lovable stands out as my favorite, despite its limitations.
For initial prototyping and proof-of-concept work, Lovable delivers real value. It excels at creating that first working version of your idea, helping you validate concepts before investing heavily in development. This approach has actually changed how I think about approaching new products—starting with an AI-generated foundation that can be refined later.
However, I strongly disagree with claims that tools like Lovable eliminate the need for developers. This is simply untrue. While you might get a simple landing page or basic functionality working, anything more complex quickly becomes problematic without development knowledge. Without coding experience, users won’t know how to debug issues, understand security implications, or even formulate effective prompts when things go wrong.
Looking ahead, I believe we’re still years away from AI coding agents fully replacing human developers, especially for projects with any meaningful complexity. Tools like Lovable are evolving rapidly, but they remain assistants rather than replacements—valuable for specific use cases when their limitations are properly understood.