Build with confidence.
A practical guide to the builder, the AI, and what happens after export.
From prompt to source
Tell the AI what you need. It receives your current files and recent conversation, returns structured multi-file edits, and validates syntax and static references before applying them. Failed edits leave the previous source intact.
Every iteration has a checkpoint
Undo, redo, or restore up to 20 versions. Saved account projects include files, chat, checkpoints, and model preferences. Signed-out work lives in this session only.
AI that actually runs
The default calls the built-in no-key service. Local Ollama and CORS-enabled compatible servers send real inference requests to your endpoint. Run the coding connection test in model settings; never infer availability from a provider label.
Preview is not production hosting
Static HTML/CSS/JavaScript runs in a sandboxed local frame. React/TypeScript previews use Sandpack’s external browser bundler, which receives the preview source and may need network access. Do not preview secrets or confidential source with this external bundler. Server code, database migrations, and deployment scripts are source files, not running services in this browser.
Publishing & deployment
Choose the hosting setup that fits your generated project.
- Export the complete source.Use Export to download a ZIP. For a React project, install the dependencies in package.json, run its build command, and review the output and tests locally.
- Deploy to your chosen host.Static hosts can serve HTML/CSS/JS or a built React output. Use your hosting provider’s upload or Git deployment workflow. Free-tier limits and commercial-use rules vary; check the provider before deploying.
- Configure production services.Backend logic needs a runtime. Authentication, databases, uploads, notification delivery, and payments need their own service configuration. Store credentials as backend environment variables, never in client code or exported public HTML.
- Verify before announcing success.Check the real HTTPS URL, routes, forms, backend responses, logs, and production permissions. A successful export or asset upload alone does not verify a production application. Custom domains and certificates are configured with your host and DNS provider.
Know the boundaries
Source generation is real; services are not silently provisioned.
The builder cannot execute arbitrary backend servers, install operating-system dependencies, or provision custom domains from this browser. It generates source and setup instructions, and supports static preview and React browser bundling. For a full production stack, export and deploy to a compatible runtime. Public preview links publish a read-back-verified snapshot, not a provisioned backend. Updating a published project replaces that snapshot only when you publish again; Unpublish removes the public link.
The generated website runs in an isolated frame without access to your account SDK, private project storage, or credentials. Integrations in model settings connect your builder account; they do not automatically configure every exported website. Ask the AI for the integration source and configure the website’s own credentials.
Visual editing targets static HTML and synchronizes changes to index.html. Dynamic React elements must be changed through code or AI. Runtime errors appear in build logs; “Ask AI to fix” submits those real errors with your current files.
Common questions
Is the AI free and unlimited?
No paid key is needed for built-in AI, but the free service permits 100 requests per hour per app, 8,192 output tokens, and a 60-second timeout. Your self-hosted model has no builder-imposed message quota; hardware, hosting, and your provider still impose practical limits.
Can I choose an open-source coding model?
Yes. Connect Ollama in model settings, choose a model identifier, and test a real response. The preset is Qwen3-Coder 30B. Model recommendations are not promises about your hardware. The built-in model’s identity and license are undisclosed.
How does failure recovery work?
Malformed AI responses and validation failures do not replace your source files. Retry with a smaller request or switch models. Model settings can enable an explicit built-in fallback. Restore any retained checkpoint through History.
Are my projects private?
Signed-in project data uses the server-enforced account scope in the built-in key-value store. Uploaded assets are public. Exported files and published HTML are public once you share or host them. Do not put secrets in source code.
Can I build payments, login, or a backend?
Ask the AI to generate the appropriate implementation and setup documentation for your chosen stack. Those services must be configured and deployed separately; the preview does not simulate a successful payment, user account, email, or database write.