Graffiticode Site de Engajamento de Agentes
graffiticode/www
Projeção viva do MCP server com descoberta de agentes e artifacts executáveis
Ver no GitHubSite canonical que funciona como projeção viva do servidor MCP Graffiticode, oferecendo descoberta de agentes, catálogo de ferramentas gerado dinamicamente e superfícies machine-readable (llms.txt, .well-known/mcp.json) sempre sincronizadas com o servidor real. Primeira parada para developers e agentes IA conectarem ao Graffiticode e executarem chamadas bem-sucedidas.
Instalação
npx convez-ai graffiticode/www1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Descoberta e onboarding de agentes IA no Graffiticode
- Catálogo dinâmico de linguagens e ferramentas disponíveis
- Validação de endpoints e contrato machine-readable para integração
Veredito

Abordagem sólida de single source of truth: todos os documentos derivam do contrato TypeScript, garantindo que llms.txt, catálogo e configuração de agentes nunca divergem da realidade. Link checking em CI força conformidade estrita entre documentação e serviço vivo.
README
A tradução para o português deste README ainda está na fila.
graffiticode.org (www)
The canonical agent-engagement site for Graffiticode — a projection of the live MCP server.
Why this exists
This is where an AI agent — or the developer behind it — lands to discover Graffiticode, connect to the
MCP server, and make a first successful call. It leads with a working artifact, a correct no-credential
quickstart, a tool catalog generated from the real language registry, and machine-readable discovery
surfaces (llms.txt, .well-known/mcp.json) that stay in lockstep with what the server actually
exposes. Everything an agent reads here is something it can act on and get back a real result.
Design principles (enforced, not aspirational)
- Single source of truth — everything derives from
data/contract.ts+data/languages.json. The homepage,/agents,/languages,llms.txt, and.well-known/mcp.jsonare all projections of it, so they cannot disagree. - No credential in any instruction — the free plan is the no-
Authorizationpath; the install configs carry no bearer token. - Documented == real —
npm run check:links(run in CI) fetches every URL the discovery files emit and fails on any 404. Broken promises can't ship. - Artifact-first — the homepage and each
/languages/[id]page embed a live Graffiticode item (see "Showcase items" below), not a screenshot.
Develop
npm install
npm run dev # regenerates discovery files, then next dev
npm run build # prebuild generates discovery files, then next build
Key scripts
| Command | What it does |
|---|---|
npm run generate | Write public/llms.txt + public/.well-known/mcp.json from the contract (runs automatically on prebuild; offline). |
npm run sync:languages | Refresh data/languages.json from the live MCP server's list_languages, preserving curated fields. On-demand. |
npm run check:links | "Documented == real" gate. BASE_URL=http://localhost:PORT npm run check:links to verify a running build. |
npm run verify | generate + check:links. |
Showcase items (live embeds)
Each language can carry a showcaseItemId in data/languages.json — a real item created via the MCP
server, embedded as living proof. When it's null, the embed shows an honest "coming online"
placeholder. To populate: create items with create_item, then drop the returned ids into
data/languages.json (no code change needed).
Deploy
Cloud Run, mirroring the existing site's setup but as a separate www service (no auto-cutover of the
graffiticode.org DNS):
npm run gcp:build # Cloud Build (Docker) → gcr.io/graffiticode-app/www
npm run gcp:deploy # deploy to Cloud Run service `www` (us-central1)
SITE_URL (build arg / env) drives absolute URLs in the discovery files, robots, and sitemap.
Stack
Next.js 15 (app router) · React 19 · TypeScript · Tailwind v4. No MDX/docs-template chrome.