convez.ai
MCP ServerIntegração & APIsIntermediário
qiun

WebMCP Airline

qiun/webmcp-airline

Demo MCP pronto para agentes: reserva de voos multi-cidade end-to-end

Ver no GitHub
InicialPluginScore · de 0 a 100

Um servidor MCP que simula um site de reserva de voos multi-cidade totalmente funcional. Implementado com Vite + React + TypeScript, expõe 7 ferramentas para que agentes IA possam pesquisar aeroportos, buscar voos, selecionar trechos, informar passageiros e realizar compras com confirmação humana.

Instalação

$npx convez-ai qiun/webmcp-airline

1 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

Estrelas
0
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Testar agentes IA em jornadas complexas de e-commerce
  • Demonstrar capacidades de navegação e automação de UI para agentes
  • Validar implementações do padrão WebMCP em cenários realistas

Veredito

Excelente referência prática para quem desenvolve agentes MCP. O código é limpo, bem documentado e cobre um fluxo de negócio real (busca multi-cidade, seleção, dados de passageiros, compra com confirmação). Ideal para testar interoperabilidade e validar implementações do padrão WebMCP.

README

A tradução para o português deste README ainda está na fila.

webmcp-airline

▶ Live demo: https://webmcp-airline.vercel.app

An airline — flight search and (mock) ticket "purchase" working for humans and exposed to AI agents through WebMCP. Supports one-way, round-trip, and multi-city. Every human action has an equivalent as a tool operating on the same logic (src/store/booking.ts).

Why this exists

This is one of four open demo sites built specifically to be agent-ready via WebMCP. We publish them so that anyone building AI browser agents — or working on the WebMCP standard and the MCP-B polyfill itself — has realistic, end-to-end targets to test against, not toy pages. The full booking journey here (search multi-city flights, pick a flight per segment, enter passengers, buy the ticket) can be completed by a human in the UI and driven entirely by an AI agent through WebMCP tools, with human-in-the-loop confirmation on the purchase.

The goal is simple: make it easy to verify that an agent can really operate a WebMCP-enabled site, end to end. Fork it, point your agent at it, and see what works.

WebMCP demo set — live demos (source): acme · shop · stays · airline

Stack

  • Vite + React 18 + TypeScript (strict), react-router-dom
  • Tailwind CSS, lucide-react
  • zod + zod-to-json-schema
  • MCP-B polyfill (@mcp-b/global); no backend — booking stored in localStorage (webmcp-airline-booking)

Running

pnpm install
pnpm dev        # http://localhost:5173
pnpm build
pnpm lint
pnpm typecheck

WebMCP tools (7)

Toolread-onlydestructivehuman-in-loop
search_airports✓––
search_flights✓––
get_flight_details✓––
select_flight–––
set_passenger_details–––
get_booking_summary✓––
book_flight–✓YES

Passenger and pricing model

interface PassengerCounts { adults: number; children: number; infants: number; }
// Rules: adults>=1; infants<=adults; total<=9.

Price per adult comes from the flight generator; Child = 75%, Infant = 10% of the adult price. The total price is computed for the whole party with a breakdown by passenger type and summed across segments. Dates are validated (format, not in the past). book_flight is registered only when every segment has a selected flight and complete passenger details; it requires human confirmation (confirmWithUser → <ConfirmDialog/>) and returns a PNR.

Mock data

  • src/data/airports.ts: 32 airports { iata, city, country }.
  • src/data/flights.ts: a deterministic flight generator for (from, to, date, cabin) (seed from a hash of the arguments — the same inputs = the same results, for repeatable agent tests).

How to test with an agent

  1. Native Chrome 146+ with the WebMCP flag (chrome://flags).
  2. MCP-B polyfill (via the adapter) + @mcp-b/webmcp-local-relay bridge → Hermes / OpenClaw / your own MCP agent.
  3. Smoke test:
    pnpm dev
    npx @mcp-b/webmcp-local-relay
    MCP_URL=http://localhost:8080/mcp node examples/agent-smoke-test.mjs
    
    Goes through search_flights (multi-city, 2 adults + 1 child) → select_flight ×N → set_passenger_details → book_flight.

API stability (June 2026)

WebMCP is a W3C Community Group Draft Report. API changes → you fix only src/lib/webmcp.ts. Changelog: developer.chrome.com/docs/ai/webmcp.

Deploy

Vercel / Netlify / GitHub Pages. SPA fallback to index.html. HTTPS required. Set Content-Type: application/json for /.well-known/webmcp.

License

MIT — see LICENSE.

Ver o original no GitHub

Bom para

Relacionados