Build your app with AI: the honest guide

Type “build my app with AI” into a search box and four different kinds of product answer. A chat window that turns a description into a working web app. A coding agent that sits inside a developer’s editor. A no-code platform with a drag-and-drop screen builder. And a service where you describe what you want and a team builds it and then keeps it running. They are all real, they all work, and they are not for the same person at the same moment.

This guide is written by people who sell one of the four. Idea to Production is a described-and-operated service, and we say so where it matters below, because a guide that hides its interest is not much of a guide. What follows is meant to survive being read by someone who uses Lovable or Replit every day and finishes it thinking the description was fair.

Four things that all answer to “build my app with AI”

Prompt-to-app builders

These are the products most people mean. You open a chat, describe an application, and the tool writes the code, runs it, and publishes it to a URL — often within the first session. The current field, each described in its own terms:

• Lovable builds and deploys web applications from natural language, with a built-in backend and hosting. It is for a founder or product person who wants a working prototype on a URL today and is comfortable running it themselves, or handing the repository to a developer.
• Replit Agent builds full-stack apps from plain language and publishes them on Replit’s own cloud, with a managed database, file storage and sign-in built in. It is for someone who wants to build, publish and run in one place and is comfortable with credit-based usage billing.
• Bolt.new runs in the browser, writes JavaScript web and mobile apps from prompts, and publishes them on Bolt Cloud hosting with a built-in database. It is for a designer, founder or product person who wants to prototype and publish quickly, with a token budget they control.
• v0, from Vercel, builds full-stack Next.js apps from chat, previews them in a sandbox and publishes them to a Vercel project. It is for someone who is comfortable in the Vercel dashboard for domains, hosting and alerts, and has or will have a developer to take the repository forward.
• Base44, owned by Wix, writes a React front end on its own built-in backend — database, sign-in, file storage, functions and hosting — from chat. It is for someone who wants a working internal tool or customer app with a database and sign-in in one place, without choosing a hosting or database provider.
• Emergent writes a React or Next.js front end, a Python FastAPI back end and a MongoDB database from chat, or an Expo mobile app, and deploys it on its own hosting. It is for someone who wants a full-stack web or mobile app with its own back end on a URL within a session.
• Rocket.new generates a Next.js web app or a Flutter mobile app from a prompt, a Figma file or a template, on a Supabase back end, and publishes the web app to Netlify. It is for someone who wants web and mobile from the same tool and is happy to run their own Supabase project.
• Databutton, now called Riff, writes a React and TypeScript front end, a Python FastAPI back end and a PostgreSQL database from chat and hosts the app. It is now aimed at procurement and operations teams in manufacturing, with ERP-connected agents, an audit trail and EU data residency.

What they cost: most have a free tier, and a first paid plan sits between roughly US$20 and US$30 a month, less if billed annually. Usage — the AI’s work, and often the hosting, database and AI features inside the published app — is measured in credits or tokens on top of that. Riff does not publish a price list; onboarding is by contacting them.

Who they are genuinely for: anyone who wants to see an idea on a screen today, and who is prepared to be the person who runs it afterwards, or who has a developer to hand it to.

Coding agents

Claude Code, Cursor and Codex are a different category. They write, edit and run code, but they live in a developer’s editor or terminal and assume someone who can read what they produce, decide whether it is right, and take responsibility for deploying it. They are sold as a monthly subscription per developer, with usage on top for heavier plans.

They are a developer’s tools, not a founder’s. If you have an engineer, they make that engineer faster. If you do not, they hand you a repository and a set of decisions that need an engineer to make.

No-code platforms

Bubble, Softr, Glide and Retool predate the current wave. Instead of writing code, you assemble screens, workflows and data tables in a visual editor, and the platform hosts the result. Pricing is per application, per editor or per user, and the bill grows with the team that uses it.

They are for someone who wants a tool their own team will use — an internal dashboard, a portal for a known set of users, a workflow that replaces a spreadsheet — and who is comfortable with the application living inside the platform for as long as it exists. An application you sell to a stranger’s IT department raises questions — data model, permissions, uptime — that are yours to answer on any of them.

Described-and-operated services

This is what Idea to Production is, so read this paragraph as the vendor talking. You describe the application in plain English. A team builds it, with real accounts, roles and permissions from the first version, on a backed-up database, on your own domain. Then the same team keeps it running: watching it from outside, applying patches on a schedule, testing the restores, and answering when it breaks. It costs a licence plus the resources the application actually uses — no salary, no equity — and a first working version takes days rather than minutes, because it has real accounts and real data from the first version. The source is yours.

It is for a founder or business whose application has to work in front of a paying customer, and who does not want to employ engineers to keep it alive. It is not the right first stop if you are still finding out whether anyone wants the thing. For that, a builder is faster and cheaper, and we would tell you so on a call.

What the self-serve builders are genuinely good at

Speed. Seven of the eight builders above publish a working version within the first session, and Lovable, Replit Agent, Bolt.new, v0 and Base44 do it on the free plan. A founder can open a chat at nine in the morning and send a link to a prospect by lunch.

Price. A first paid plan costs about what a team lunch costs. For an idea that may not survive contact with a customer, that is the right size of commitment. Spending six figures on an agency build to find out whether people want something is a trade the builders have made unnecessary.

Learning by doing. Describing an application and watching it appear teaches you what you actually want faster than any specification document. The screen you asked for turns out to be wrong; you see it, you say so, it changes. By the fifth iteration you understand your own product in a way weeks of meetings would not have reached.

Code in hand. Lovable, Replit Agent, Bolt.new and v0 sync to a GitHub repository — a copy of the code under your own account — and Emergent and Rocket.new do so on paid plans. StackBlitz states the code Bolt.new produces is yours for any legal purpose; Vercel states it does not own what v0 generates from your prompts; Emergent states you own everything it generates and may host it anywhere. Rocket.new puts the database in your own Supabase project from the first day. If you later leave, you leave with something.

A clickable prototype in a customer conversation. A prospect who clicks through a live screen tells you things a slide deck never draws out. They reach for the button that is not there. They ask what happens next. A working prototype turns a hypothetical conversation into a concrete one, and concrete conversations are where companies get founded.

For the stage it is designed for, a self-serve builder is the right answer, and most of the people reading this should use one before they talk to anyone like us.

What none of them do

Once an application exists, about nine jobs begin, and none of them ends. We wrote them up in full in The cost of keeping software alive; briefly:

1. Watching. Something has to check the application is answering, from outside, on a short interval, and tell a person when it stops. If nobody does it, you learn about the outage from your best customer.
2. Patching. Every library the application depends on will eventually carry an advisory. If nobody applies the patches, the application that was safe last year is quietly a liability this year.
3. Restoring. Backups are common; tested restores are rare. If nobody rehearses the restore, you find out whether the backup works on the one day it has to.
4. Deploying safely. Changes need a staging environment, a check suite and a rollback path. If nobody builds them, every change is a small gamble made in production.
5. Scaling. Capacity has to follow demand, and demand has to be measured. If nobody watches it, you find your ceiling on the day you get written about.
6. Permissions. Who can see whose records, and who can change them. If nobody designs this, the first customer with two staff members finds that one of them can see everything.
7. The security questionnaire. Every serious customer sends one. If nobody holds the answers, the deal stalls in procurement while you write them from scratch.
8. The incident at 2 a.m. When something breaks, someone has to contain it, fix it, and explain afterwards what changed. If nobody is on call, the fix waits until you wake up, and the customer does not.
9. Somebody accountable. One person whose job it is when the application fails. If nobody holds that job, you do, in evenings you did not budget.

To be exact about it: some of the builders provide pieces of this. Lovable takes a daily database backup and keeps about fourteen days of them, and its project monitoring, in beta, emails the owner about errors it finds. Replit Agent and Emergent run uptime checks on the published app, and Replit Agent emails you when it is down. Hosting on Lovable, Replit Agent and v0 scales with traffic. Base44 gives app users admin and user roles. Most run a security scan that lists findings, some prepare a fix for you to review, and Bolt.new applies most of the fixes and lists the rest. What none of them does is hold the job. No vendor documents a rehearsed restore; the scan runs, but the person who reads it, applies the fix and republishes is you. Four of the eight state that their support does not debug the code in your project, and the rest do not document a position — a reasonable line for a self-serve product to draw. It just means the nine jobs are yours.

Choose by where you are, not by the tool

The question has a different right answer at each stage, and most bad decisions come from answering it for a stage you have not reached yet.

An idea and a window

You have a problem you think is worth solving and a few weeks before something else claims your attention. Use a builder. Ship the prototype to a URL. Put it in front of ten people who have the problem, and listen more than you demonstrate. Do not pay anyone to make it robust, because you do not yet know what it is. The job at this stage is learning, and the builders are the fastest way to learn.

A first paying customer

Someone has agreed to pay. This is the wall. The questions change from “can you show me” to “where is my data”, “what happens if it goes down”, “who else can see this”, and “is there a backup”. None of them is about building; all of them are about running. This is the moment to be honest about whether you are going to hold the nine jobs yourself, hire someone to hold them, or pay a service to hold them. We wrote a checklist for exactly this moment in From AI prototype to production, and it applies whichever way you choose.

A first enterprise deal

A larger company wants to buy, and with the purchase order comes requirements the prototype never had. Single sign-on with their identity provider. An audit log of who did what. A security questionnaire from a procurement team that will not sign until it is answered. At this stage the question is not which tool you used but who can say yes to that list and be responsible for the answers. Our own position is on the compliance page, which says what is in progress rather than claiming everything is done.

If you have already built something

If you are reading this with a Lovable or Replit Agent app already running, that is not a problem to apologise for. It is the best brief anyone could give us.

A prototype shows exactly which screens, flows, tables and roles you want. We treat it as the brief rather than as code we inherit, and rebuild from it with real accounts, a backed-up database and monitoring from the start, because that is faster and safer than vouching for generated code we did not write. What you learned from the prototype is not lost; it is the whole point.

If you want to see how each builder compares on the after-the-build questions, side by side and sourced from the vendors’ own documentation, the comparison pages are the place. If you would rather talk about where you are, contact us and describe what you have built. The prototype’s URL is a fine place to start.

The honest summary

Stage	Right tool	What it costs you	What to watch for
Testing whether anyone wants it	A prompt-to-app builder	A free tier or about US$20 to US$30 a month, plus your evenings	The demo is not the product; do not promise anyone it is
A prototype and interested prospects	The same builder, with the repository synced to GitHub	Credits or tokens that climb with use	The data model and permissions you improvise now are hard to undo later
A first paying customer	A described-and-operated service, or an engineer you employ	A licence plus usage, or a salary	Who is told when it goes down, and how quickly
A first enterprise deal	Whoever can answer the questionnaire and be held to it	SSO, audit logs and compliance evidence, kept current	The questionnaire arrives before the contract does
A tool for your own team only	A no-code platform, or a builder	Per-seat or per-application pricing that grows with the team	The person who built it leaves, and nobody else can change it

Whichever row you are on, building has become the cheap part. The expensive part is everything after, and it is not what the builders sell. If you reach the row where someone has to hold that job, our promises are written down as five guarantees, each with what it covers and what it does not: security, compliance, reliability, scalability and efficiency. Read them the way you would read a vendor’s documentation, because that is what they are.