IndexedAI — Claude Code Skill para Consulta MCP
guidodr27/claude-skill-indexedai
Consulte sites via MCP sob demanda, zero overhead de contexto
Ver no GitHubSkill que consulta websites via seus próprios servidores MCP, retornando dados estruturados sem overhead de contexto. Carrega MCP sob demanda quando invocado, nunca no startup, e oferece fallback automático para WebFetch se o MCP não estiver disponível.
Instalação
npx convez-ai guidodr27/claude-skill-indexedai1 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
- Consultar documentações técnicas estruturadas (Stripe, GitHub, OpenAI)
- Buscar informações de políticas de retry e limites de taxa em APIs
- Consultar múltiplos domínios em paralelo sem desperdício de tokens
Veredito

Solução elegante que resolve um problema real do MCP — overhead de contexto desnecessário no startup. A implementação é sólida com fallback gracioso, mas depende de sites que exponham endpoints MCP discovery (ou do serviço IndexedAI.tech). Recomendado para quem trabalha frequentemente com documentações de APIs e quer eficiência de tokens.
README
A tradução para o português deste README ainda está na fila.
indexedai — Claude Code Skill
Query any website via MCP — structured data, zero context overhead.
The problem with normal MCP servers
When you add an MCP server to Claude, it loads every tool definition at startup — filling your context window before you've typed a single message.
/indexedai works differently. It loads MCP on demand, only when you invoke it. No idle overhead, no wasted tokens.
What it does
/indexedai queries any website through MCP (Model Context Protocol) and returns clean, structured data — not raw HTML.
It works with sites that expose a /.well-known/mcp.json discovery endpoint, including sites indexed by IndexedAI.tech — a service that generates hosted MCP endpoints and llms.txt files for any website.
If no MCP server is found, it falls back to plain WebFetch automatically.
User: /indexedai docs.stripe.com how does webhook retry work?
Claude: Checking for MCP server at docs.stripe.com…
[MCP: docs.stripe.com]
Stripe retries failed webhook deliveries up to 3 days
with exponential backoff starting at 1 hour...
Key benefits
| Benefit | Detail |
|---|---|
| Zero context overhead | MCP loads only when invoked — never at startup |
| Structured data | Semantic content via MCP, not raw HTML scraping |
| Always returns an answer | Graceful WebFetch fallback when MCP is unavailable |
| Parallel queries | Multiple domains discovered and queried concurrently |
How it works
/indexedai <domain> <question>
│
▼
GET /.well-known/mcp.json
│
┌───┴────────────────────┐
│ MCP endpoint found │ No MCP / error
▼ ▼
Initialize (JSON-RPC) WebFetch fallback
List tools [Web: domain.com]
Call best-matching tool
[MCP: domain.com]
Discovery happens via .well-known/mcp.json — the MCP standard for advertising server endpoints. If a site doesn't have one natively, IndexedAI.tech can generate one.
Install
One command:
mkdir -p ~/.claude/skills/indexedai && \
curl -fsSL https://raw.githubusercontent.com/guidodr27/claude-skill-indexedai/main/SKILL.md \
> ~/.claude/skills/indexedai/SKILL.md
Manual:
- Download
SKILL.mdfrom this repo - Place it at
~/.claude/skills/indexedai/SKILL.md - Restart Claude Code (or reload skills)
Usage
# Full URL with question
/indexedai https://docs.stripe.com "webhook retry policy"
# Domain shorthand
/indexedai openai.com rate limits
# Multiple sites queried in parallel
/indexedai stripe.com and docs.github.com webhook events
Output format
| Prefix | Meaning |
|---|---|
[MCP: domain.com] | Structured data via MCP endpoint |
[Web: domain.com] | Fallback — plain WebFetch |
Want MCP for your site?
IndexedAI.tech analyzes any website and generates:
- A hosted MCP server endpoint (queryable via this skill)
- A custom
llms.txtfile for AI agent readability - An Agent Readiness Score across 5 dimensions
- Analytics — see which AI systems (Claude, GPT, Perplexity, Cursor…) access your site, how often, and what they query
Requirements
- Claude Code (any version supporting skills)
- No additional dependencies
Examples
See the examples/ directory for real-world usage walkthroughs.
Contributing
See CONTRIBUTING.md.
License
MIT © 2026 guidodr27