Editor

Augmented vision. Unlimited potential.

Auvira.ai is an AI website teammate for small businesses — describe what you want, and Auvira creates, edits, remembers, and deploys your site. No drag-and-drop. No code.

This repo is Auvira's editor — the Next.js app owners use to clone, build, chat-edit, preview, and publish.

Dev note: npm package / repo folder name remains ai-website-migration-agent for tooling compatibility.

Every time an owner works on their site, this app:

• Extracts business facts from an existing URL (clone) or a short intake form (scratch)
• Generates a one-page Next.js + Tailwind site from deterministic templates — the LLM produces JSON (siteConfig, copy, colors), not React source
• Validates every build locally before GitLab commit or Vercel deploy
• Edits via a GitLab-backed chat agent — scoped targets, rollback, diff review
• Remembers — calls the Auvira.ai backend for coaching hints, intent grounding, and Phoenix traces
• Analyzes — per-project session quality dashboard (keyword cloud, coaching context, turn history)
• Publishes — syncs workspace to GitLab, SHA-pins Vercel deploy, marks the site live

The Auvira.ai backend runs the brain's memory. This repo runs the editor and LLM.

Built with Next.js 14 · TypeScript · MongoDB · GitLab · Vercel · Gemini · Auvira.ai Backend API

Website builders generate pages. They don't tell you whether the AI understood the owner, whether the site still builds, or what "do the same" actually means.

Auvira's editor closes that gap:

Turn every owner message into a scoped, build-verified edit — then make the next one smarter.

Deep dive: docs/EDIT_AGENT.md · docs/OBSERVABILITY.md

Production editor deploys via Vercel from main. Set NEXT_PUBLIC_APP_URL to your production domain. Root / redirects to /intro.

Owner message in editor chat
       ↓
① GET  /context          → coaching hints from prior edits
①b POST /intent          → grounded intent sentence
②      Edit agent runs locally (plan → tools → verify)
③ POST /turns            → score + Mongo + Phoenix trace
       ↓
Repeat — each session gets smarter

POST /turns  →  grade + Phoenix trace (builder.turn)
       ↓
GET /context  →  coaching hints  →  Gemini planner  →  POST /turns  →  …

Observability is on by default (shipped read-only client key). Set OBSERVABILITY_ENABLED=0 to opt out.

Integrating or extending the edit loop? → docs/EDIT_AGENT.md · docs/OBSERVABILITY.md

flowchart TB
  subgraph ui ["Auvira.ai Editor (this repo)"]
    dash[Dashboard]
    clone[Clone / Scratch agents]
    editor[Chat + Preview]
    agent[Edit Agent]
    analytics[Analytics dashboard]
  end

  subgraph build ["Build pipeline"]
    tmpl[Builder templates]
    gate[Build gate]
  end

  subgraph external ["External services"]
    gitlab[(GitLab repos)]
    vercel[Vercel deploy]
    mongo[(MongoDB)]
    backend[Auvira.ai Backend API]
    phx[Arize Phoenix]
  end

  dash --> clone
  dash --> editor
  editor --> agent
  editor --> analytics
  clone --> tmpl
  agent --> tmpl
  tmpl --> gate
  gate --> gitlab
  gitlab --> vercel
  agent --> backend
  clone --> backend
  analytics --> backend
  backend --> phx
  editor --> mongo
  clone --> mongo

buildEditContext → resolveEditTarget → assessEditAmbiguity → resolveImplicitReferences
  → planEdit → executePlan → domain tools → verify → summarize

• HTTP: POST /api/projects/[projectId]/code-agent/edit/stream (SSE)
• Runner: src/lib/project-workspace/websiteEditRunner.ts
• Agent: src/lib/project-workspace/edit-agent/index.ts

Owners drag a section or element from the preview iframe onto chat to pin a selectedTarget for scoped edits.

• Product catalog (Phase 2) — MongoDB Product model; sync into draft siteConfig; static export with inquiry CTAs (no native checkout)
• Site Manager — scheduled checks (uptime, phone, hours, services, banner expiry); incidents and fix proposals via /api/projects/[projectId]/site-manager/*
• Site analytics (generated sites) — optional runtime on customer sites; collect endpoint at /api/analytics/collect
• Commerce (Phase 3, not enabled) — Stripe checkout stub in src/lib/commerce/stripeConfig.ts; not wired in .env.example

Primary actions avoid GitLab/Vercel jargon — backup copy, draft preview, live website.

Auth: Google OAuth (Auth.js) · GitLab required for owner edits (409 without link)

Full edit-agent map: docs/EDIT_AGENT.md

Prerequisites: Node.js 20 · MongoDB · Google OAuth · GitLab token + group · Gemini API key

git clone https://github.com/athan37/auvira.ai.git
cd auvira.ai
npm install
cp .env.example .env.local
cp .env.example .env          # optional; recommended for npm test
npm run dev

Open http://localhost:3000. Sign in → dashboard → create or open a project.

See .env.example for the full list.

1. GitLab personal access token: scopes api, read_user, read_repository, write_repository
2. Google Cloud Console → OAuth 2.0 Client ID (Web)
3. Redirect URI: http://localhost:3000/api/auth/callback/google

Deploy: Connect repo to Vercel · production branch main · set env vars (MongoDB, OAuth, GitLab, Gemini). Add production URL to Google OAuth redirect URIs.

Customer sites: Publish syncs workspace → GitLab → SHA-pinned Vercel build. Verify with:

npx tsx scripts/e2e-deploy-verify.ts --project-id <id>

Test:

npm run test:all
npm run typecheck && npm run build   # when types or build paths change

CI (.github/workflows/ci.yml): typecheck → test → test:contracts → build — no live LLM.

For agent workflow details and preview drag-to-chat troubleshooting, see AGENTS.md.

Auvira.ai — the AI operating layer for small businesses.
This repo powers the editor behind every edit.