← projects

Project case study

Red Thread

live

Agentic operating system for luxury hospitality. Live A2A handoff between two real machines (Threadkeeper on Vercel ↔ elias on a Mac mini via Cloudflare tunnel), real Claude on each side.

Open full ↗
platform · 2026 · repo ↗ · live ↗
AnthropicElevenLabsA2ACloudflare TunnelVercel
Event context Cerebral Valley × Rosewood Hospitality 2030 Hackathon Built at Rosewood Sand Hill, Menlo Park · May 2026 1st place · Rosewood stay

Red Thread is an agentic operating system for luxury hospitality: a sense of place, threaded through every guest. It won 1st place at the Cerebral Valley x Rosewood Hospitality 2030 hackathon. Live demo at app.redthread.boutique.

What it does

A single-URL dashboard where you submit a reservation, then watch the agent work. Begin briefing runs a 9-step live chain: CRM cross-property lookup, three web-research steps (LinkedIn, X, press) gated by the guest’s Hold-the-Thread privacy score, an AviationStack flight lookup, Claude reasoning for luggage / customs / transit, and an ETA composition. Each card streams reasoning tokens and expands on click to show the exact input, result, tech provenance, and timing. The brief renders into a Dossier across three editorial zones, and a Brief Me button reads it aloud through ElevenLabs.

Flip Sand Hill to Hong Kong in the header and the same guest produces a completely different dossier. That is the no-cookie-cutter moment: the system holds a sense of place, not a template.

Architecture

The front end is Next.js on Vercel, streaming everything over SSE. The distinctive piece is the agent-to-agent handoff: hit Agent handoff and Threadkeeper (the Rosewood-side agent, on Vercel) opens a live conversation with elias (my personal agent, running on a Mac mini in my house, exposed through a Cloudflare tunnel at elias.redthread.boutique). Real Claude runs on each side, a structured manifest is extracted from the negotiation, and the dashboard rerenders with the agreed preferences. The A2A contract is published at /.well-known/agent.json so any agent can discover how to talk to it.

Implementation notes

  • Anthropic (Claude Opus + Haiku). Two-tier by design: Haiku 4.5 streams the fast per-step reasoning and the privacy “Discretion” ladder; Opus composes the final dossier. The lesson was that the discretion ladder had to be implemented twice in lockstep, a fast fixture path for slider repaints and the live Claude path, so the UI never burns a model call just to move a dial.
  • ElevenLabs. Convai for the live voice intake (a 5-question call with the Alice persona) plus a TTS proxy for the brief read-aloud. What was innovative for demo-day: a scripted two-voice fallback that streams a fixed Alice + Eric conversation when live Convai is flaky, so the demo never dies on stage.
  • AviationStack + live web tools. Real flight lookups and web research, cached per query because the live tools burn 10 to 30 seconds each; the cache is what made the live chain demo-able in real time.
  • Vercel + Cloudflare Tunnel. The tunnel is the whole trick behind the A2A demo: a consumer Mac mini at home becomes a real, addressable second agent on the public internet, talking to a cloud agent as a peer.

Brand and craft were part of the build, not bolted on after: cream and ink, the red Chinese endless knot 紅線 as the brand mark, the thread of fate.