PantryPilot: an Alexa+ MCP server that fights food waste

Track: Alexa+ (self-hosted MCP server, Streamable HTTP) | Mini challenge: AWS Builder (Amazon Bedrock)

"Alexa, good morning."
"Add six eggs that expire in ten days."
"What's expiring soon?"
"What can I cook tonight?"
"Plan my dinners for the week."
"Which medicines expire this month?"

PantryPilot gives Alexa+ a persistent household pantry plus recipes that use the soonest-expiring food first — reducing food waste one voice command at a time. Recipes come from Amazon Bedrock, with a built-in recipe book (offline_recipes.py) when Bedrock is unavailable.

Note: our AWS account is currently blocked from calling Bedrock models (FR-09), so the live demo uses the offline recipe book. GET /health shows the last Bedrock result.

How it works

flowchart LR
    user([User speaks]) --> alexa["Alexa+<br/>or the web simulator"]
    alexa -- "MCP over HTTPS<br/>Streamable HTTP, /mcp" --> server["PantryPilot MCP server<br/>FastMCP, 20 tools<br/>AWS App Runner"]
    server --> db[("SQLite: pantry, shopping list,<br/>meal plan, medicine cabinet")]
    server -- "Converse API" --> bedrock["Amazon Bedrock<br/>Nova Micro"]
    server -. "if Bedrock is unavailable" .-> book["Offline recipe book"]
    ecr["Amazon ECR"] -- "container image" --> server

Alexa+ reads each tool's description to decide which one to call, sends the call over MCP, and speaks the short reply. Every reply is written to be heard: one to four sentences, no lists or markdown.

Related MCP server: Chef's Brain

Tools (all voice-first: short, speakable, no markdown)

Medicine cabinet: inventory only, never medical advice

Families keep half-forgotten medicines for years. PantryPilot tracks them like food: expiry dates as printed on the pack ("03/2027", "March 2027"), oldest batch used first, expired stock flagged and never "used", and safe disposal guidance (pharmacy take-back). Questions about dosing, safety, or which medicine to take get a fixed reply pointing to the label, a pharmacist, or a doctor; the same guard keeps medical questions out of the Bedrock recipe prompt.

Quick start (local)

# 1. Clone and set up
git clone https://github.com/abhisuman2001/pantrypilot-alexa-mcp.git
cd pantrypilot-alexa-mcp

python -m venv .venv
.venv\Scripts\activate          # Windows
# source .venv/bin/activate     # macOS / Linux

pip install -r requirements.txt

# 2. Configure AWS credentials (needed for Bedrock)
cp .env.example .env            # then fill in your values
# AWS_ACCESS_KEY_ID=...
# AWS_SECRET_ACCESS_KEY=...
# AWS_REGION=us-east-1
# BEDROCK_MODEL_ID=amazon.nova-micro-v1:0   (enable in Bedrock console first)
# MCP_API_KEY=                              # optional: leave blank for local dev

# 3. Run
.venv\Scripts\python server.py    # http://localhost:8000/mcp

The server starts with CORS enabled so the web simulator works out of the box.

Web Simulator (demo asset)

The server hosts the simulator at /. Start the server, then open:

http://localhost:8000/                                  # local
https://kykh233phz.us-east-1.awsapprunner.com/          # live, no setup needed

The page calls /mcp on the same server it was loaded from. Serving it over http(s) (not file://) is also what lets the browser microphone work.

Features:

• Alexa-style voice chat UI (dark mode, animated ring)

• One-click quick actions for the pantry, shopping list, and recipe tools

• Forms for Add Item / Use Item / Add to Shopping List

• 🎤 Web Speech API voice input (Chrome/Edge)

• Natural language parsing: "add 6 eggs expire 10 days", "what can I cook?", etc.

• Settings panel: configure server URL + API key

• Optional studio-quality voice from ElevenLabs (see below); otherwise the best browser voice

Optional: ElevenLabs voice

Set ELEVENLABS_API_KEY on the server and the simulator speaks replies with an ElevenLabs voice. The page posts the reply text to the server's /tts endpoint, and the server calls ElevenLabs, so the key never reaches the browser. Without a key, or if a call fails, the simulator falls back to the browser voice.

To protect the account's credits on a public server, /tts accepts at most 400 characters per request and 20 requests per visitor every 10 minutes, caches repeated phrases, and stops after ELEVENLABS_CHAR_BUDGET characters (8000 by default, reset on restart).

This only changes the simulator. Real Alexa+ reads PantryPilot's replies in Alexa's own voice.

Use it from any MCP client (the way Alexa+ calls it)

The live server is a standard MCP endpoint, so any MCP client can discover and call its tools.

MCP Inspector (no install):

npx -y @modelcontextprotocol/inspector
# Transport: Streamable HTTP, URL: https://kykh233phz.us-east-1.awsapprunner.com/mcp
# Then: List Tools, and call daily_briefing or add_medicine

Claude Code, so an AI agent picks the tools from their descriptions, as Alexa+ does:

claude mcp add --transport http pantrypilot https://kykh233phz.us-east-1.awsapprunner.com/mcp
# then ask: "add six eggs that expire in ten days, then tell me what to cook tonight"

Raw JSON-RPC with curl:

curl -s https://kykh233phz.us-east-1.awsapprunner.com/mcp \
  -H "Content-Type: application/json" -H "Accept: application/json, text/event-stream" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"daily_briefing","arguments":{}}}'

Local development works the same with http://localhost:8000/mcp.

Run tests

.venv\Scripts\pytest -v
# 168 passed

Smoke-test a deployed server end to end (it also reports whether Bedrock answered or the offline recipes were used):

python scripts/live_smoke.py                                   # live App Runner URL
python scripts/live_smoke.py http://localhost:8000             # or any other base URL

Environment variables

Privacy & Terms

Served inline by the App Runner deployment:

• Privacy Policy: https://kykh233phz.us-east-1.awsapprunner.com/privacy

• Terms of Use: https://kykh233phz.us-east-1.awsapprunner.com/terms

• Health check: https://kykh233phz.us-east-1.awsapprunner.com/health

Deploy (Docker → AWS App Runner)

# One-command deploy (builds, pushes to ECR, creates App Runner service):
Set-ExecutionPolicy -Scope Process -ExecutionPolicy Bypass
.\deploy.ps1

Live deployment:

• Service URL: https://kykh233phz.us-east-1.awsapprunner.com

• MCP endpoint: https://kykh233phz.us-east-1.awsapprunner.com/mcp

• ECR image: 448049796441.dkr.ecr.us-east-1.amazonaws.com/pantrypilot:latest

Alexa+ Registration

TODO(verify): registration steps pending — see track Resources page.
The MCP endpoint to register is: https://<your-domain>/mcp

Submission checklist

• All 20 tools callable via MCP Inspector and web simulator

• pytest passes (168 tests)

• Bedrock called in code in suggest_recipe and weekly_meal_plan (preference-aware, offline recipe book fallback)

• Bedrock answering live — blocked at the AWS account level, see FR-09

• Bearer-token auth (MCP_API_KEY), input validation, no stack traces exposed

• Deployed over HTTPS — https://kykh233phz.us-east-1.awsapprunner.com/mcp

• Privacy policy and Terms of Use served at /privacy and /terms

• set_preferences, schedule_meal, get_meal_schedule added (Step 9)

• Medicine cabinet: 6 tools, inventory only, medical-advice guard

• daily_briefing for an Alexa+ morning routine

• Redeploy to App Runner with new tools

• Demo video recorded (≤ 3 min, voice demo first) — record and link here

• Devpost description written — see SUBMISSION.md

• Product feedback for every tool/API/SDK used — see SUBMISSION.md

• Friction log completed (+10% bonus) — see FRICTION_LOG.md (9 items, severity + workaround + suggestion each)

Demo note: the live deployment runs with MCP_API_KEY unset so judges can call it without a key. The pantry is shared demo data and resets on every redeploy.