Barrio

One Alexa+ add-on for every shop in the neighbourhood.

Comment

Inspiration

In my town, Écija (Spain), my barber takes bookings on his phone between haircuts, and my dad answers the phone at his Tex-Mex restaurant during the dinner rush. They'll never hire a developer — and they'll certainly never get an Alexa+ add-on through a store listing and certification. I wanted the smallest possible step between "a business with a phone number" and "a business Alexa+ can talk to".

What it does

The owner fills in a short form — or takes a photo of the paper menu — and Barrio gives the business its own MCP server. Customers ask Alexa+ about prices and dishes, check free times and book or cancel by voice; on an Echo Show they see the business's own cards (free times they can tap, dish details with clear gluten/allergen notes, the booking confirmation). Bookings appear in the owner's dashboard and in the calendar on their phone.

And because no small business will certify its own add-on, Barrio also exposes a directory endpoint: one MCP server — one Alexa+ add-on — in front of every business on the platform. "Alexa, ask Barrio for a haircut at La Familia tomorrow."

It's seeded with the published services, menu and prices of two real businesses in Écija, used with their owners' permission — Restaurante El Rancho, my dad's restaurant, and Peluquería La Familia, the barber I go to (I built his website). It's a demo: bookings made here aren't sent to them.

How I built it

  • Following the Alexa+ MCP Toolkit docs: Streamable HTTP (2025-11-25; also negotiates 2025-06-18 and 2025-03-26, which Alexa+'s client sends), outputSchema + structuredContent on every tool, "still working" messages, tool annotations, ~5–10 ms server time with a read-through cache and region pinning, optional OAuth 2.1 with RFC 8414/9728 metadata, privacy/terms pages, and every store asset at Amazon's sizes — generated from code.
  • A real MCP client in the simulator: the simulated Alexa+ (a Claude agent) connects to the business's public endpoint over Streamable HTTP — initialize, tools/list, tools/call, resources/read — exactly as the track's FAQ describes for the simulated path. Every reply shows its MCP calls.
  • MCP Apps for Echo Show: a ~26 KB self-contained card (768×480 base canvas, Amazon's card tokens, light/dark, English or Spanish from the host locale, fitted to Echo Show 8 and 15) linked from every tool. When a question could mean two things, the card shows the options to tap. The simulator and a public "screens" page render it with the official MCP Apps host, so judges can see exactly what an Echo Show would show.
  • Designed for voice: "tomorrow", "el próximo viernes", "5:30pm", "half past five"; "anything free after five?" narrowed on the server, so the model never filters a long list of times itself; English aliases for Spanish services ("beard trim" → Arreglo de barba); ambiguous requests ("ensalada", "tacos with guacamole") ask instead of guessing; references without look-alike characters; alternatives on both sides of a lunch break; a retried request never books twice.
  • No double bookings: a Postgres function with a per-business advisory lock re-checks capacity (the peak number of tables in use), the per-phone limit and an hourly ceiling against booking floods.
  • Honest about allergens: gluten-free only when the restaurant says so; undeclared allergens → "ask the staff". Menu photo import never extracts allergens.
  • Owner tools: photo/text menu import with review, split-shift hours, calendar subscription, email alerts, customer confirmation emails with a self-cancel link, 12-month data retention.

Challenges

  • Finding the real Alexa+ requirements (they're in the toolkit docs, not the hackathon page) and rebuilding against them.
  • Getting the MCP Apps view from 520 KB to ~26 KB without breaking compatibility with the official host.
  • Things that only break against the real world: the current Sonnet model rejects a forced tool choice (the menu import had to move to structured outputs), Supabase returns at most 1,000 rows per request, and a US evening is already tomorrow in Spain.
  • Stateless Streamable HTTP on serverless, and booking rules that survive real life: lunch breaks, closing time, DST, 100 people asking for the same slot.

Accomplishments

  • 335 automated tests + a real-Postgres concurrency test (21 checks) + 71 browser checks in CI, plus 11 Supabase integration tests: protocol compliance, 100-customer races, 3,000 fuzzed calls with hostile input (zero 5xx).
  • Three review rounds — the last one split by area (protocol, security, database, AI integration, Echo Show design, Amazon's docs). Findings like "tacos with guacamole" getting the guacamole's gluten-free answer, or a visitor's test business named "La Familia" receiving the real shop's bookings; every one fixed with a regression test.
  • Live deployment, measured from Spain: sequential MCP round trips p50 150 ms / p95 222 ms (well under the 500 ms budget Alexa+ sets for its strictest add-on category); 2,000 calls with 100 concurrent clients, 0 errors. Locally: p50 6 ms.

What I learned

MCP turns "integrate with a voice assistant" into "describe your business well". Most of the work was making tool descriptions, errors and cards useful to a model talking to a person — and making sure it never makes things up.

What's next

Text-message confirmations, staff-level availability (several barbers with their own schedules), Google Business Profile import, and onboarding more businesses in Écija.

Built With

Share this project:

Updates

Submission history