Roasi

Your website is getting roasted tonight.

An AI agent that audits your site with the energy of a Filipino senior dev — four beers deep, completely out of patience, and brutally correct.

---

Meet Roasi

Roasi is the AI agent at the heart of this product. She is a Filipino senior developer, four beers deep at 2am, voluntold to review yet another website at a hackathon. Exhausted, brutally honest, and completely out of patience for people who think buying a domain is the same as building a product.

Her voice never changes. Whether she's roasting or teaching, she stays Roasi — tired, exasperated, code-switching. She explains LCP like she cannot believe she has to. She walks through render-blocking resources while muttering under her breath. The teaching is real. The warmth is not.

Every line she writes is data-driven. Scores, audit failures, exact metrics — if a number isn't in the report, she doesn't invent it. Vague roasts are for people who are also bad at their jobs.

Her code-switching is structural. Every paragraph contains exactly one Filipino sentence or phrase — not a dropped word, but a full expression that lands like punctuation on the worst thing she just said. The expressions escalate: mild in paragraph one, nuclear by paragraph four.

The roast is four paragraphs, no headers:

1. React to what the site is trying to be — start mid-thought
2. Specifics — exact audit failures, exact scores, exact metrics
3. Pull back — who built this and what do they think they built
4. Find the one thing that almost worked, name it sincerely, then dismantle it

Roasi does not modify code, write files, or touch the user's codebase. She scans, analyzes, teaches, and guides.

---

How It Works

User submits a URL
       │
       ▼
  ┌─────────────────────────────────────────────┐
  │              Roasi Agent Loop               │
  │                                             │
  │  1. Scan    →  Unlighthouse audit           │
  │  2. Analyze →  Priority-tier categorization │
  │  3. Roast   →  Brutal, structured critique  │
  │  4. Guide   →  Actionable recommendations   │
  └─────────────────────────────────────────────┘
       │
       ▼
  Results stream in real time via SSE

---

Tech Stack

Monorepo

	Tool	Purpose
	Turborepo	Task orchestration, build caching
	pnpm workspaces	Package management
	Biome	Formatting + linting
	TypeScript	Shared configs via packages/typescript-config

Frontend

	Library	Version	Role
	Next.js (App Router)	16.1.6	Web framework with Turbopack
	React	19.2	UI runtime
	Tailwind CSS	4.1	Styling
	shadcn/ui	base-lyra	Component primitives
	Pixi.js	8.13	2D sprite animations (Roasi mascot)
	Motion	12.38	UI animations

AI / Agent

	Library	Role
	Vercel AI SDK (ai 6.x)	Agent loop, streaming, tool use
	@ai-sdk/mistral	LLM provider (mistral-small-latest)
	Unlighthouse	Headless Lighthouse auditing engine
	Puppeteer	Browser automation for scans
	Zod	Tool input schema validation

Infrastructure

	Service	Role
	Vercel	Hosting + serverless functions
	Supabase	Persistent storage — scrape cache, AI response cache, chat memory
	Upstash Redis	Rate limiting (IP / host / global tiers)
	useworkflow.dev	Durable workflow orchestration
	Cloudflare Turnstile	Bot protection (optional)
	Firecrawl	Web scraping for scan tools

---

Architecture

Monorepo Structure

/
├── apps/
│   ├── web/               # Next.js 16 main application
│   └── design-system/     # Vite-based component explorer
├── packages/
│   ├── ai/                # LLM agent, tools, prompts
│   ├── ui/                # Shared React components (shadcn/ui)
│   ├── sprite-animations/ # Pixi.js Roasi mascot animations
│   ├── eslint-config/     # Shared ESLint rules
│   └── typescript-config/ # Shared tsconfig presets

Agent Tool Loop

POST /api/roast
  └─ ToolLoopAgent (Vercel AI SDK)
      ├─ planTool / updateStep    — declare & track workflow steps
      ├─ scanSite                 — Unlighthouse audit (4 modes)
      ├─ analyzeScanReport        — priority-tier categorization
      ├─ memoryTool               — persistent core/notes/conversation storage
      └─ loadSkill / bash         — dynamic skill execution

Roasi's Prompts

Roasi's behavior is defined in prompt files under packages/ai/src/agents/roasi/prompts/:

File	Purpose
PERSONALITY.md	Voice, tone, code-switching rules, roast structure, profanity guidelines
AGENT.md	Workflow, tools, audience calibration, output format, content guard

Streaming

Responses stream via SSE (Server-Sent Events). The client uses the Vercel AI SDK useChat hook, and tool results render with custom per-type components as they arrive — no batching.

Storage

Supabase is the primary persistent store:

Table	Contents
scrape_cache	Scraped site data and roast metrics
ai_response_cache	LLM response cache (deduplication middleware)
chat memory	Conversation history via memory tools

Filesystem (.workspace/ locally, /tmp in production) handles ephemeral state:

.workspace/
├── .memory/
│   ├── core.md                     # Facts injected into every agent turn
│   ├── notes.md                    # Archival notes
│   └── conversations/
│       ├── {chatId}.jsonl          # Message history (JSONL per chat)
│       └── titles/{chatId}.txt     # Chat titles for sidebar
└── .reports/{timestamp}/           # Lighthouse HTML reports (served via /api/reports)

Scan Modes

Mode	Pages	Description
default	1	Homepage only
targeted	Custom	Specific paths
smart	~20	Sitemap-guided
full	200	Full crawl

SSRF protection blocks all private/local hosts and cloud metadata endpoints.

Priority Tiers

Tier	Audits
🔴 1 — Critical	is-crawlable, meta-description, LCP, CLS, TTI
🟠 2 — High impact	color-contrast, image-alt, WebP images, render-blocking, unused JS/CSS
🟡 3 — Enhancements	Everything else

---

Getting Started

Prerequisites

• Node.js >= 20
• pnpm 9.x

Setup

pnpm install

Copy .env.example to apps/web/.env and fill in your values:

# App
NEXT_PUBLIC_APP_URL=http://localhost:3000

# AI
MISTRAL_API_KEY=your_key_here

# Web scraping
FIRECRAWL_API_KEY=your_key_here

# Supabase (scrape cache, AI response cache, chat memory)
SUPABASE_URL=your_supabase_url
SUPABASE_SERVICE_ROLE_KEY=your_service_role_key
SUPABASE_DB_PASSWORD=your_db_password

# Upstash Redis (rate limiting)
UPSTASH_REDIS_REST_URL=your_redis_url
UPSTASH_REDIS_REST_TOKEN=your_redis_token

# Bot protection (Cloudflare Turnstile) — optional
# NEXT_PUBLIC_TURNSTILE_SITE_KEY=your_site_key_here
# TURNSTILE_SECRET_KEY=your_secret_key_here

Dev

pnpm dev

Opens at http://localhost:3000.

Build

pnpm build

---

UI Components

shadcn components are added to packages/ui:

pnpm dlx shadcn@latest add button -c apps/web

Import from the shared package:

import { Button } from "@roaster/ui/components/button";