Inspiration
Our demo shop is Fade & Co., a fictional one-chair barbershop in Dalston, London (demo data). Marcus runs it alone. He is the barber, the receptionist, the cleaner and the accountant.
His bookings arrive the way they do for most small shops: Instagram DMs, WhatsApp, texts, the odd phone call. On a normal day that's around 40 messages, and almost all of them land while he has clippers in his hand. He can't stop a skin fade halfway through to reply to "any chance of 4 o'clock?", so the reply goes out an hour later, by which point the customer has booked somewhere else.
Two more things cost him money every week:
- No-shows. About one in five Saturday appointments doesn't turn up. Saturday is his busiest and most valuable day, and an empty chair on a Saturday can't be sold again.
- Same-day gaps. When someone cancels at 11am for a 3pm slot, there are usually people who would have taken it. Marcus just has no time to text round and find them.
None of this needs a big system. It needs someone to answer "is there a time?" straight away, with the real diary in front of them. Alexa+ can now hold a conversation and call tools on a business's behalf, so we built the tools.
What it does
A customer says, "Alexa, can I get a skin fade at Fade & Co. tomorrow afternoon?" The assistant checks Marcus's real diary and offers up to three free times. The customer picks one and gives their name and mobile number. The booking is made and a confirmation text goes out. Marcus gets a one-line note: "New booking by voice. Nothing for you to do."
- Real availability. Free times come from opening hours, the lunch break, existing bookings and the length of the service. A 60-minute cut-and-beard won't be offered at 18:15 when the shop closes at 19:00.
- Saturday deposit. Saturday bookings carry a £5 deposit with a pay link in the confirmation, the rule Marcus chose to cut Saturday no-shows.
- Reschedule and cancel by voice. The customer gives their mobile number and the assistant finds their booking. Rescheduling moves it; it never leaves a duplicate behind.
- Waitlist refill. When a day is full, the assistant offers the waitlist. When someone cancels, every person on that day's waitlist gets a text straight away: "3pm on Saturday just opened. First to reply YES gets it."
- Owner's briefing. Marcus can ask, "What does my day look like?" and hear his appointments in order, the day's takings and the waitlist size.
- Helpful errors. If a time is taken, the answer is a sentence the assistant can say out loud, with the closest free times.
All of this is exposed as a standard MCP server, so Alexa+ can use it, and so can any other MCP client.
How we built it
- MCP server on Cloudflare Workers. The
/mcpendpoint uses the official@modelcontextprotocol/sdkwith the Streamable HTTP transport, targeting spec 2025-11-25. It runs stateless: every request builds a fresh server and transport, so any Worker instance can answer any request. - Nine tools, one catalogue. Each tool is defined once with zod: input schema, output schema, description and annotations (
readOnlyHinton lookups,destructiveHinton cancel). Every tool returnsstructuredContentmatching itsoutputSchema. - A resource and a prompt.
business://fade-and-co/infogives an agent the services, hours, deposit rule and today's date in one read; thebook_haircutprompt opens a booking conversation with the receptionist rules attached. - Booking engine. A 15-minute grid, breaks, service length, same-day buffer, deposits by weekday, phone matching, the waitlist cascade and the owner's notes. Data lives in Cloudflare D1, with times stored in the business's timezone.
- Agent Skill.
skills/book-appointment/SKILL.mdpackages the conversation rules: offer at most three times, collect name and mobile before booking, never read out IDs, confirm before cancelling. - An Alexa+-style simulator. Without an Alexa+ developer device, the web page acts as one: tap the ring and talk (Web Speech API) or type. A small agent loop on Workers AI (Llama 3.3 70B, function calling) calls the same tool code the MCP server uses, and a side panel shows every tool call, the day's diary and every text sent.
- Tests. 34 tests run without a Cloudflare account: in-memory SQLite behind a D1-compatible adapter, a frozen clock, and the official MCP client talking to the Worker in-process.
Challenges we ran into
- Testing without a device. We built the simulator so the tools could be exercised by a model we didn't control, with a system prompt close to how a voice assistant should behave.
- Making answers sound right. A JSON list of 30 free times is correct but useless when spoken. Voice rules went into tool descriptions, error messages written as sentences, and the Agent Skill.
- Dates and "now". "Tomorrow afternoon" has to become a London date, not UTC. The engine returns spoken dates ("Thursday 8 October") alongside ISO dates, so the assistant never formats a date itself.
- Bugs the tests found. A blank service name matched the first service; rescheduling 15 minutes later was blocked by the old booking; after closing time the day was called "fully booked"; a customer could join the same waitlist twice. All fixed and covered by tests.
Accomplishments that we're proud of
- The whole booking loop works by voice: ask, book, reschedule, cancel, waitlist, owner briefing.
- A cancellation turns into a text to every waiting customer in the same request.
- One tool catalogue drives the MCP server, the simulator's agent and the tests, so what we demo is what Alexa+ gets.
- It runs on the Cloudflare free tier. A small shop could run this for nothing.
What we learned
- For voice, the tool's output is part of the conversation design: spoken dates, short lists and errors that already contain the alternative.
outputSchemaandstructuredContentcatch drift between code and description.- Stateless MCP is a good default on serverless: keep the state in the database, not the connection.
- Small businesses don't want a dashboard. Marcus's version of success is not having to look at his phone between cuts.
What's next
- Publish the integration for Alexa+ and test it on real devices with real customers.
- Real SMS or WhatsApp messages and real deposits through a payment link.
- "First to reply YES" handled end to end.
- Onboarding for other one-person businesses (nail techs, tattoo artists, dog groomers, physios) from a short form.
- Per-business sign-in for the owner tools.
Built With
- agent-skills
- alexa+
- cloudflare-d1
- cloudflare-workers
- javascript
- llama-3.3
- mcp-sdk
- model-context-protocol
- node.js
- web-speech-api
- workers-ai
- zod