Week 2: The first build sprint

Last week we committed to the thesis. This week we built the first version.

The stack decision

Next.js 15 with App Router, TypeScript strict, Supabase for auth and database, Tailwind for styling. No exotic choices — speed comes from not deliberating.

The critical bet: build the MCP server from day one, not as an afterthought. If Corexi's thesis is "live where the agent lives," the MCP surface can't be a phase-two add-on.

Day by day

Monday-Tuesday: Auth, database schema, RLS policies, basic dashboard shell. Scan pipeline skeleton — Playwright captures a screenshot, sends it to Claude Vision, parses structured findings.

Wednesday-Thursday: axe-core integration for accessibility verification. Every vision-LLM finding gets cross-checked against axe. This is the "verified WCAG" claim — not vibes, not "the AI said so," but deterministic a11y rules confirming the AI's assessment.

Friday: MCP server with the first three tools: get_ux_rules, review_code, list_sites. Tested with Cursor — the agent could already ask Corexi for UX rules and review a code snippet.

Weekend: Design token support. Upload your CSS variables or Tailwind theme, and fix_code outputs var(--your-token) instead of hardcoded values. The more tokens you upload, the more tailored every fix becomes to your design system.

What we deliberately cut

• No billing. Free tier only, enforce limits in code, wire Stripe later.
• No multi-language. English only, next-intl infrastructure in place but no translations.
• No onboarding wizard. Add URL, hit scan, read findings.
• No marketing site. A single landing page with "what it does" and "sign up."

The lesson

Scope is a weapon. Every feature you don't build is a day you ship earlier. We launched with 3 MCP tools, not 9. The other 6 came in the following weeks, informed by what the first users actually asked for.