Wibe — AI Website Builder
Describe an app in plain English — an AI coding agent builds a working Next.js app in a cloud sandbox, with live preview and code.
(01)Overview
Wibe is an AI SaaS for founders, non-developers, and developers who want working prototypes fast. You describe what you want; an autonomous coding agent spins up an isolated cloud sandbox, writes and installs everything a real Next.js + shadcn/ui app needs, and hands back a live preview alongside the source. Keep chatting and the agent edits the existing project surgically instead of starting over — with every version saved, credit-based plans, and authentication built in.
(02)What we built
- 1. Prompt to production-shaped app — a complete Next.js application with shadcn/ui and Tailwind, generated inside an isolated E2B sandbox.
- 2. Conversational iteration — follow-up messages switch the agent into modification mode, so it reads existing files and changes only what's needed.
- 3. Live preview of the running sandbox in an embedded browser — refresh, copy the URL, or open it in a new tab.
- 4. Code view with a file tree and syntax highlighting for every generated file.
- 5. Version history — each AI response stores its own snapshot, and clicking it switches both preview and code to that version.
- 6. Credit-based Free and Pro plans (Pro: 100 credits a month) with Clerk billing, a live usage indicator, and upgrade prompts.
- 7. Clerk authentication with route protection — every project, message, and fragment is scoped to its owner.
- 8. One-click starters — Netflix, admin dashboard, kanban, file manager, YouTube, store, Airbnb, and Spotify clones.
- 9. AI-written project titles and friendly progress replies alongside each build.
- 10. Polished workspace — resizable chat/preview split, light and dark themes, Lottie loading states, and a WebGL aurora background.
(04)How it works
Generation pipeline
- 01PromptA Clerk-protected tRPC mutation validates the prompt, spends one credit, and records the message.
- 02Background jobAn Inngest event starts a durable, retryable job — the request returns instantly while the agent works.
- 03Cloud sandboxE2B boots an isolated microVM from a custom Next.js template with shadcn/ui pre-installed and pre-compiled.
- 04Coding agentA GPT-4.1 agent loops through terminal, write-file, and read-file tools — up to 15 iterations — until it reports the task complete.
- 05Preview & saveThe sandbox URL and merged files are saved as a versioned fragment in Postgres; a lighter model writes the title and reply.
(05)Engineering highlights
Token-efficient context
Instead of sending whole codebases back to the model, the agent receives the file list with line counts plus the last five messages, and reads files on demand.
Surgical edits, not rewrites
Existing projects are wrapped in a modification-mode brief that enforces read-before-edit and minimal changes; old and new files are merged when saved.
Durable agent runs
Every tool call runs as an Inngest step, so long generations survive failures, retry safely, and stay observable.
Warm previews
The custom sandbox image boots Next.js and polls until the first page compiles, so the preview is ready the moment the agent finishes.
Fair-use credits
A Postgres-backed rate limiter enforces plan quotas over a rolling 30-day window without any extra infrastructure.
Engineered system prompt
A 467-line system prompt defines the agent's environment, new-versus-existing project detection, and a step-by-step editing workflow.
(06)Inside the build
Screenshots from the running system — select any image to enlarge.
More work
View all →
Want something like this?
Tell us about your product or platform — we'll get back to you within one business day.