convez.ai
MCP Server
MyTrip-ai

octo-mcp-server

MyTrip-ai/octo-mcp-server

Ver no GitHub
InicialPluginScore · de 0 a 100

Reference MCP server for the OCTO standard (tours, activities & attractions) — agentic booking gateway: one server, many OCTO suppliers.

Instalação

$npx convez-ai MyTrip-ai/octo-mcp-server

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
1
Forks
0
Acompanhando
1
Saúde
37%
Último commit
há 1 semana
Versão
—
Issues abertas
0
Contribuidores
1

README

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

OCTO MCP Server — unofficial demo

⚠️ Unofficial · not affiliated with, or endorsed by, OCTO. This is an independent demo exploring how AI agents could use the open OCTO standard for tours, activities & attractions. It books against mock suppliers (and, optionally, a Ventrata test supplier). No real bookings, no real charges.

What happens when an AI can actually book a tour — not just chat about one? This is a tiny Model Context Protocol server that lets an AI agent discover and book in-destination experiences across every connected OCTO supplier through one server — safely, with a human approving before anything is charged.

It's the working answer to a strategic question for OCTO: as travelers start asking their AI assistants to book things, who are the rails between the agent and the supplier?


▶ Watch the 20-second demo

npx asciinema play media/octo-demo.cast

Prefer to read it? media/octo-demo.txt is the full transcript.

Try it yourself (zero install)

npx github:MyTrip-ai/octo-mcp-server

That launches the Guided Concierge — a non-technical, numbers-and-Enter walkthrough: search → check availability → hold → approve → confirm, with the OCTO machinery shown as it happens. Runs fully offline on bundled demo suppliers; no API key required.

Requires Node 20+. First run builds automatically.

Three modes, one server

CommandWho it's forWhat it does
(none)anyoneGuided Concierge — the interactive booking walkthrough
connectyour AIWires this MCP server into your Claude Desktop / Claude Code / Cursor, then hands off — your assistant gains the booking tools
demoa roomAn on-rails narrated pitch that auto-plays the whole story
npx github:MyTrip-ai/octo-mcp-server connect   # give your own AI booking superpowers
npx github:MyTrip-ai/octo-mcp-server demo       # the narrated pitch

Why this matters for OCTO

  1. A new distribution channel. When travelers ask their AI to book experiences, OCTO can be the rails between agent and supplier.
  2. Build once, reach every supplier. A single MCP server fronts all OCTO suppliers — the same standardization that powers today's resellers powers AI agents tomorrow. Add another OCTO supplier (Bókun, Ventrata, …) and it appears with zero new code.
  3. OCTO can lead this. A canonical "agent-ready" profile — an octo/mcp or octo/ai-content capability — would let suppliers advertise AI-readiness and put OCTO ahead of the curve. (An open question this demo is meant to start, not settle.)

How it works (the design that makes it safe)

This is deliberately not a 1:1 wrapper of the OCTO REST API — that's a footgun for LLMs. See DESIGN.md for the full rationale. The essentials:

  • One server, many suppliers — every supplier sits behind one OctoSupplierAdapter (mock or live HTTP). One search reaches them all.
  • The model never touches opaque IDs — it works with friendly handles; the OCTO availabilityId and idempotency uuid stay server-side, so an agent can't double-book.
  • Money is always human-readable — OCTO's integer + currencyPrecision is normalized to "$180.00 USD"; raw integers never reach the model.
  • Human-in-the-loop on spend — confirm_booking refuses without explicit approval and is flagged destructive. The AI proposes; a human disposes.
  • Capabilities + auth negotiated server-side — the model never manages headers or keys.

Use live OCTO inventory (optional)

The Guided Concierge's onboarding offers to take a free Ventrata test key for live data; or set it in a git-ignored .env:

VENTRATA_OCTO_ENDPOINT=https://api.ventrata.com/octo
VENTRATA_OCTO_API_KEY=your-free-test-key
# optional: ANTHROPIC_API_KEY=...  (upgrades the chat brain from deterministic to Claude)

Get a free key at https://dashboard.ventrata.com/octo/signup (test supplier "EdinExplore").

Architecture

src/
  index.ts              the OCTO MCP server (stdio) — 9 tools, resources, prompts
  server.ts             assembles suppliers (mock + optional live Ventrata)
  octo/                 OCTO types + adapter seam (mockAdapter, httpAdapter)
  chat/engine.ts        shared chat engine (deterministic + Claude brains, parsers)
  cli/                  the CLI: ui · concierge · connect · demo
scripts/                bridge (web chat) + smoke/live/cli/connect/demo tests + make-cast
web/                    a browser chat over the same MCP server (npm run bridge)
media/                  the recorded demo (.cast + .txt)

Develop & verify

npm install
npm run cli            # the Guided Concierge
npm run smoke          # MCP server, mock suppliers (12 checks)
npm run live           # real OCTO round-trip via Ventrata (needs a key)
npm run cli-smoke      # guided flow, end to end
npm run connect-smoke  # config writer + handshake
npm run demo-smoke     # narrated pitch
npm run cast           # regenerate media/octo-demo.{cast,txt}
npm run bridge         # browser chat at http://localhost:8787

License

MIT. Built as a reference artifact for the conversation about AI agents and the OCTO standard. Again: unofficial — not affiliated with OCTO.

Ver o original no GitHub

Relacionados