Lowtide

An Alexa+ MCP server: ask Alexa when to run the dishwasher, the car or the hot water. Lowtide plans it around tonight's live electricity prices and grid carbon, and draws it on the Echo Show.

Comment

For the judges (60 seconds)

Links: Echo Show demo · Website · Code · Alexa+ fit table · Friction log (16) · Countertop · Upstream fix: mcp-handler#200

  1. Open https://d1soo7510nfzf4.cloudfront.net/echo. No sign-in; a demo household in London is made for you.
  2. Say or type "When should I run the dishwasher? It needs to be done by seven", then "Yes please", then "How much have I saved?"
  3. Watch the Conversation panel: every MCP tool call Amazon Nova Micro makes on Bedrock, live, with Amazon Polly speaking the answer.
  • Live data, three markets: Octopus Agile in all 14 British regions (any postcode or city), ComEd hourly in Illinois, EPEX day-ahead in Germany, plus the grid's carbon forecast. No mocks.
  • Alexa+ ready: an MCP server (2026-07-28 and the 2025 revisions), nine voice-first tools, an MCP App view for the Echo Show screen, elicitation, OAuth 2.1 for Alexa+ account linking.
  • Runs on AWS: Lambda behind CloudFront in Sydney, Nova Micro on Bedrock, Polly, S3, EventBridge Scheduler planning tonight for every household, CloudWatch. No access keys: each function signs with its own role.
  • Proof: 147 tests and CI, a live smoke test of every tool, 16 friction-log entries.

A real conversation on a simulated Echo Show: Amazon Nova Micro on Bedrock picks Lowtide's MCP tools; the screen is its MCP App view.

Inspiration

Millions of homes now pay a different price for electricity every half hour: Octopus Agile in Britain, ComEd Hourly Pricing in Illinois, and dynamic tariffs in Germany, which every supplier has had to offer since 1 January 2025. On 15 October in London a unit cost 19p in the afternoon and 55p at six in the evening. The money is in the big loads: charging a car at the cheapest time that day cost £4.25 instead of £7.30, and heating a hot-water cylinder £1.16 instead of £1.97. On peaky days like that one, two habits are worth about £610 a year. But nobody reads a price table while loading the dishwasher. "When should I run it?" is a question you ask out loud, with your hands full, in the kitchen. That is exactly where Alexa lives.

What it does

  • You ask Alexa when to run the dishwasher, the washing, the tumble dryer, the car charger or the immersion heater, with a deadline ("done by seven", "by tomorrow morning"), a start ("not before 10pm") or a goal ("greenest, please"). A deadline it can't make gets a straight "it can't be done by then" and the earliest finish.
  • Lowtide prices every start that finishes in time and answers twice: one sentence to say, with the cost and the saving, and an MCP App view for the Echo Show screen, the next 24 hours on a dial with the run marked.
  • Say yes and it saves the run for the household, then counts what you've saved. "Is now a good time to put the tumble dryer on?" gets a straight answer.
  • Tonight's plan, before anyone asks: after tomorrow's prices are published, Amazon EventBridge Scheduler plans the coming night for every household, and the Echo shows it as a waiting notification.
  • Savings are quoted against a typical start, never against the evening peak. The car alone moves 7.3 kW of load out of the evening.
  • The same server works in Claude, ChatGPT and VS Code: add the guest endpoint, or sign in with OAuth and pick a household.

Tonight's prices as a tide: high water is dear, low tide is cheapest, and the display says when to start the dishwasher.

How we built it

  • MCP server: Next.js route handlers with mcp-handler and the MCP TypeScript SDK v2, stateless Streamable HTTP, serving MCP 2026-07-28 natively and 2025-11-25, 2025-06-18 and 2025-03-26 through initialize. Nine tools with voice-first results, all four annotations and Alexa+'s invoking/invoked strings.
  • MCP App view: one ui:// resource with an inline layout for chat hosts and a full-screen device layout for the Echo Show.
  • Alexa+ simulation: the Echo Show simulator is a real MCP host in the browser. Each turn goes to Amazon Nova Micro on Bedrock (Converse API, SigV4 in a few lines of node:crypto, no SDK); Nova picks the tool, the browser runs it over MCP and Polly speaks the reply. Each model turn takes 1.0 to 1.3 seconds end to end from India to Sydney.
  • On AWS: the whole Next.js app on Lambda (Node.js 22, arm64) behind a response-streaming Function URL and CloudFront, which adds a secret origin header so nothing reaches the function around it. Households, tonight's plans and OAuth token hashes in S3 with If-Match writes. Two EventBridge Scheduler runs a night; on 18 October the 16:45 run planned about 70 households. CloudWatch turns one log line per tool call into metrics, a dashboard and an alarm. One CloudFormation stack, deployed by one boto3 script.
  • Planner: every half-hour start that fits is priced across the slots it covers, with cheapest, greenest and balanced goals.
  • The website draws tonight's prices as a moonlit tide, server-rendered SVG on screen with the first paint, and shows each answer the way an appliance would: an orange LCD and a delay-start knob for each of tonight's four loads.

On AWS: one Lambda function behind CloudFront, Nova Micro on Bedrock, Polly, S3, EventBridge Scheduler and CloudWatch.

Challenges we ran into

  • Alexa+ MCP integrations are in preview for brands only, so there was no way to try Lowtide inside a real Alexa+. We built the closest honest substitute: a real MCP host with a real model on Bedrock, rendering the real MCP App view on a device-sized screen.
  • Voice changes everything: one sentence, times said the way people say them ("half seven", "by tomorrow morning"), and a screen readable across a kitchen. The exact start the model needs to save a run goes in a line marked for the assistant, never spoken.
  • AWS in an organisation: the policy blocked Bedrock API keys, IAM OIDC providers and every region but Sydney, so we signed every request with SigV4 and kept everything in ap-southeast-2.
  • Behind a Lambda Function URL, the OAuth challenge header is renamed, and CloudFront Functions don't run on error responses (friction log 15 and 16); clients discover sign-in through the well-known documents instead.
  • The nightly schedule's first guard refused the real scheduled call; a per-deploy token fixed it (friction log 14).

Accomplishments that we're proud of

  • A full voice conversation (plan, yes, saved, "is now a good time") on Amazon Nova Micro against live prices, 2.6 to 3.3 seconds from question to answer on AWS.
  • One MCP App view that works as a chat card and as a full device screen, and OAuth 2.1 for Alexa+ account linking, tested end to end with the MCP SDK's own client.
  • Countertop, an open-source bench (17 tests, CI) to test any MCP server the way a smart display would, and a fix sent upstream: vercel-labs/mcp-handler#200, an opt-in CORS option with 11 tests.
  • Honest numbers: savings against a typical start, latency measured on the live stack, and 16 friction-log entries for the Amazon teams.

What we learned

  • MCP Apps are the natural shape for Alexa+ on screens: the server owns its view, and every host renders it the same way.
  • The first line of every tool result should be something you'd be happy to hear spoken.
  • The impact is in kilowatts: the car and the hot water are what move real load.

What's next

  • Who pays: the household keeps the saving, but the natural buyer is the energy supplier, which needs load moved out of the evening peak. Britain already pays for that: the Demand Flexibility Service had 2.46 million households and businesses signed up by March 2026 and paid £943,983 for 3,917.7 MWh in winter 2024/25. With every British meter moving to half-hourly settlement by 7 May 2027, the route to market is a supplier-branded Lowtide on Alexa+.
  • Proactive notifications on a real Echo, once Alexa+ opens them to MCP servers.
  • Matter smart plugs and appliance APIs, so a confirmed plan can start the machine itself.
  • More tariffs (Octopus Go and Intelligent, PG&E, Tibber) and US carbon data.

Built With

  • agent-skills
  • alexa-plus
  • amazon-bedrock
  • amazon-cloudfront
  • amazon-cloudwatch
  • amazon-eventbridge
  • amazon-iam
  • amazon-nova
  • amazon-polly
  • amazon-web-services
  • aws-cloudformation
  • aws-lambda
  • aws-sigv4
  • carbon-intensity-api
  • mcp
  • mcp-apps
  • model-context-protocol
  • next.js
  • octopus-energy-api
  • streamable-http
  • svg
  • typescript
  • web-speech-api
Share this project:

Updates

Submission history