The "I Built This Myself" Era Just Started
A roofing company owner in Jacksonville showed us something last month that genuinely surprised us. He had built a complete job scheduling system -- customer intake form, calendar integration, crew assignment logic, automated text reminders -- using Lovable. No developer. No agency. No prior coding experience. Took him a weekend.
It was not perfect. The database schema was messy, the error handling was minimal, and it would not scale past about 50 concurrent users. But it worked. His crews were using it daily. And he built it for the cost of a monthly subscription instead of the $15,000 a custom dev shop quoted him.
This is not a novelty. This is a structural shift in who can build software -- what many are now calling vibe coding. And if you work in tech consulting, you need to understand exactly what these tools can and cannot do, because your market is changing underneath you.
A roofing company owner in Jacksonville built a complete job scheduling system -- customer intake, calendar integration, crew assignment logic, automated reminders -- using Lovable. No developer. No prior coding experience. Took him a weekend.
The New Wave: What Each Tool Actually Does
There are four tools leading this category right now, and they each have distinct strengths. We have used all of them extensively -- both for our own projects and in evaluating them for client recommendations.
Replit Agent
Replit has been an online IDE for years, but their Agent feature turned it into something fundamentally different. You describe what you want to build in plain English, and Replit Agent writes the code, sets up the database, configures the deployment, and gets a working app running -- often in minutes.
Where Replit excels is in full-stack applications. It handles both frontend and backend, sets up databases, manages environment variables, and deploys to a live URL. For someone who wants a complete web application -- not just a landing page -- Replit is currently the most capable option.
The limitation: the code it generates can be sprawling and hard to maintain. If you need to hand this off to a developer later for significant modifications, they will spend the first day just understanding the structure. It also struggles with complex state management and real-time features.
Lovable
Lovable focuses specifically on building web applications with clean, modern interfaces. It generates React code with Tailwind CSS and integrates directly with Supabase for the backend. The output is genuinely good-looking -- better default design than most developer-built MVPs.
What sets Lovable apart is iteration speed. You can describe a change, see it reflected in seconds, and keep refining. The conversation-driven development flow feels natural to non-technical users. "Make the sidebar collapsible" or "add a dark mode toggle" just work.
The limitation: it is opinionated about its stack. If you need something other than React/Supabase, Lovable is not your tool. It also hits a ceiling with complex business logic -- multi-step workflows, conditional permissions, intricate data relationships. These are doable but require increasingly specific prompting that starts to feel like programming with extra steps.
Bolt (by StackBlitz)
Bolt runs entirely in the browser using WebContainers, which means there is no server provisioning, no deployment step, no waiting. You describe what you want, and a working app appears in your browser tab. For rapid prototyping, it is the fastest tool in this category.
Bolt is particularly strong for frontend-heavy projects -- interactive dashboards, data visualization tools, single-page applications. The code quality is decent, and because everything runs in-browser, you can inspect and modify the output directly.
The limitation: anything that requires a persistent backend, user authentication, or database operations is more complicated with Bolt. It is a prototyping powerhouse but not always the best path to production.
v0 (by Vercel)
v0 takes a different approach. Rather than building complete applications, it generates individual UI components based on text descriptions or screenshots. Need a pricing table? Describe it. Want to replicate a design from a competitor's site? Screenshot it.
The output quality is exceptional -- v0 generates production-ready React components with proper TypeScript typing, accessibility attributes, and responsive design. For developers, it is a massive time-saver. For non-developers, it is a way to create professional UI elements that a developer can then integrate into a larger project.
The limitation: v0 does not build applications. It builds components. You still need to assemble them, add routing, connect a backend, and handle deployment. It is a tool for people who are already building something, not a tool for building from scratch.
This Is Not 2015 Drag-and-Drop
The critical distinction people miss: these tools generate real code. Not proprietary markup trapped in a platform. Not visual configurations that only work within one ecosystem. Actual React components. Actual Python backends. Actual SQL databases. Code you can export, version control, deploy anywhere, and hand to any developer to maintain.
Squarespace and Wix let non-developers build websites. These tools let non-developers build software. That is a categorically different thing. A website displays information. Software encodes business logic, manages data, and automates workflows.
When the roofing company owner built his scheduling system, he was not dragging blocks around a canvas. He was describing business rules in English -- "if a job is marked urgent, text the closest available crew lead" -- and the AI was writing the code to implement them. That is unprecedented.
Where They All Fall Short -- And We Learned This the Hard Way
We are not theorizing about these limitations. We got burned by them on real client projects.
Early in 2025, we tried using Replit and Lovable to build production applications for clients. Not prototypes -- actual production systems that businesses would rely on daily. The first 70% of each project went beautifully. Fast iteration, happy clients, impressive demos. Then we hit the wall.
Complex features that required tight integration between multiple systems would break in ways the browser-based AI could not diagnose. The tools would generate conflicting code across files, lose track of the project architecture, and start introducing bugs while trying to fix other bugs. We had projects that could never be completed in these environments no matter how much we prompted and re-prompted. The tools were great at getting you to 70%. Getting from 70% to production-ready was where they fell apart.
The specific failure modes are important to understand:
- Security is an afterthought. These tools generate functional code, not secure code. SQL injection prevention, proper authentication flows, input sanitization, rate limiting -- these are rarely handled correctly out of the box. For internal tools this might be acceptable. For anything customer-facing or handling sensitive data, you need a security review.
- Scale is not in the vocabulary. The generated architectures work fine for 10 or 100 users. They start breaking at 1,000. If your tool needs to handle real load, the code needs refactoring by someone who understands performance optimization.
- Browser-based tools cannot see your full codebase. This is the killer. Replit and Lovable work in sandboxed environments. They cannot navigate your existing project structure, run your test suite, or understand how a change in one file ripples through twenty others. For simple apps this does not matter. For real software it is disqualifying.
- They cannot finish complex projects. Starting is easy. Finishing is where software gets hard -- edge cases, error handling, data migration, deployment configuration, environment-specific behavior. These tools run out of runway on exactly the parts that matter most.
What We Switched To -- And Why It Changed Everything
Getting burned by browser tools led us to discover Claude Code early in 2025, and it fundamentally changed our practice. Claude Code is a CLI-based AI coding agent -- it runs in your terminal, in your actual project directory, with access to your real files, your real git history, your real test suite. We wrote a complete guide to Claude's product suite if you want the full picture.
The difference is night and day. Instead of describing your project to a browser tool that has never seen it, Claude Code reads your codebase, understands your architecture, and makes changes in context. It can run your tests, see the failures, and fix them. It can navigate a 500-file project and understand how a change in the auth module affects the API layer.
We became early adopters and have not looked back. Every production project we deliver now goes through Claude Code, not browser-based tools. The browser tools still have their place -- they are fantastic for prototyping, for non-developers exploring ideas, for getting that first 70% fast. But for finishing the job, for production-quality work, local CLI agents working in your actual codebase are the answer.
What This Means for Consulting
Our job is changing. Not disappearing -- changing. The roofing company owner did not need us to build his scheduling system. He needed us three months later when it broke under load and he wanted to add payment processing.
The consulting model is shifting from "we build it for you" to "we help you build it right." A client comes to us with a working prototype they built in Lovable over a weekend, and our job is to take that into a real development environment, audit the code, harden the security, optimize the database, and guide them on what to build next. We use Claude Code to do the heavy lifting of that transition -- taking messy AI-generated code and refactoring it into something maintainable.
Honestly, this is better for everyone. The client gets a working prototype faster and cheaper. They understand their own tool because they described every feature themselves. And when they do need expert help, the conversation starts from a working system instead of a blank requirements document. We spend less time in discovery and more time adding genuine value.
The agencies that will struggle are the ones charging $20,000 to build basic CRUD apps. That market is evaporating -- we explore this trend further in how vibe coding is disrupting SaaS. The agencies that will thrive are the ones that understand the full stack of AI tools -- browser-based for prototyping, CLI-based for production -- and know exactly when to transition from one to the other.
Our Recommendation
If you are an SMB owner with an idea for an internal tool, try building it yourself first. Start with Lovable or Replit. Spend a weekend on it. You will be shocked at what you can produce.
But know the limits. When the project starts getting complex -- when you need real authentication, real integrations, real error handling -- that is when browser tools hit their ceiling. That is when you bring in someone who works with production-grade AI development tools like Claude Code, in a real codebase, with real version control.
The barrier to building software just dropped by an order of magnitude. These browser tools are a massive part of that. Just do not mistake a great demo for a finished product. We made that mistake so you do not have to.
In fact, we learned so much from the gap between browser tools and production work that we ended up building Vibe Quest, a gamified app that teaches anyone to vibe code using terminal agents and Claude Code. The irony of building it with the exact tool we were teaching was not lost on us. Between Vibe Quest, our open-source skills library, and our interactive AI Tool Belt, we have built more free resources than most consulting firms charge for -- because we believe the fastest way to grow this market is to lower the barriers for everyone.
Browser-based tools are fantastic for prototyping, for non-developers exploring ideas, for getting that first 70% fast. But for finishing the job, for production-quality work, local CLI agents working in your actual codebase are the answer.