Introduction
A Next.js SaaS starter whose AI features get better after launch. It records every AI result, finds what fails, and proves a fix helps before you ship it.
This starter is for building a SaaS with AI features that your users can rely on. Most starters connect a model to a text box and stop there. This one records every AI result, what it cost and what the user did with it, then helps you and your coding agent fix what fails. You only ship a change once it is shown to help.
The loop
- Record. Every AI result is recorded with what it cost and how long it took.
- Measure. Your UI reports what users do with each result (kept, copied, retried, rated). The AI → Results page shows how often each AI feature worked first time, who it fails for, and what a working result really costs.
- Change. Your coding agent drafts improvements: a better prompt, a cheaper model, or a different approach altogether.
- Prove. Each change is tested on real past requests before anyone sees it, with a method built for apps with few users. What code can't judge, a person reviews blind.
- Ship safely. Each change comes to you as a pull request. Review it within the window you choose, or your agent starts it as a small trial on its own. Real users' results then decide whether everyone gets it or it is undone.
Around it you get the usual SaaS plumbing: sign-in with email links and Google, organizations, Stripe, a dollar balance for AI usage, background jobs, file uploads, docs, a blog and legal pages. Each module can be switched off.
Work with a coding agent
It comes ready for Claude Code, Codex and Cursor. Ask your agent to "add an AI feature that…" or "make this feature better", and it follows the same procedure as these docs.