Alexa, order Barrios

MCP server for the Amazon Build, Ship, Shape hackathon (Alexa+ track): "Alexa, order Barrios" lets a customer order pickup from Barrios Mexican Cantina (Oak Ridge, TN) by voice. The Alexa+ skill calls these tools; this repo is the self-hosted MCP server behind it.

What it does

Six tools:

Related MCP server: DoorDash over text

AWS integration (Builder mini-challenge)

get_recommendations calls Amazon Bedrock (Claude 3 Haiku) to suggest dishes from the real menu based on a free-text preference. The Bedrock client lives in adapters/bedrock_agent.py; core modules never import boto3. Setup:

.venv/bin/pip install boto3
export AWS_REGION=us-east-2
# credentials via standard AWS chain (env, ~/.aws/credentials, IAM role)

Each recommendation returns a real item_id you can pass straight to place_order.

Menu data is the real menu from barriosoakridge.com (scraped 2026-10-02). Prices are integers (cents). Locked facts: queso 4oz $3.50 / 8oz $7.00 / 16oz $13.00; guac 4oz $4.00 / 8oz $8.00 / 16oz $15.00.

Run locally

cd ~/workspace/latch-hackathon/alexa-order-barrios
python3 -m venv .venv
.venv/bin/pip install -r requirements.txt

# start the server (stdio transport, for MCP clients)
.venv/bin/python server.py

Orders persist to ./data/orders.db (SQLite). Override with BARRIOS_DB:

BARRIOS_DB=/tmp/test.db .venv/bin/python server.py

Transports

Streamable HTTP serves the MCP endpoint at /mcp/ (verified end to end: initialize, tools/list, tools/call all working over HTTP/SSE).

Test each tool

End-to-end through a real MCP client over stdio:

.venv/bin/python test_client.py

This asserts the acceptance case: 1x queso 8oz + 1x guac 4oz places an order totaling $11.00 with status received, and get_order_status returns it back.

With the MCP Inspector:

npx @modelcontextprotocol/inspector .venv/bin/python server.py

Then call get_menu, place_order, get_order_status from the UI.

Project layout

server.py            FastMCP server, the 3 tools
menu.py              menu data + lookup (source of truth: barriosoakridge.com)
orders.py            SQLite layer (orders table, advance_status demo helper)
adapters/aws_lambda.py  AWS boundary - NOT WIRED (see below)
test_client.py       stdio end-to-end test
requirements.txt     pinned

AWS-later plan

AWS is not funded yet, so this repo is local-first: no AWS SDK imports anywhere (grep -rniE "boto3|botocore" --include="*.py" . only matches adapters/ comments). When credits land:

1. Hosting - run the FastMCP server on Lambda (via Mangum) or ECS/Fargate; register its URL as the Alexa+ skill's MCP endpoint.

2. Bedrock - conversational layer (intent parsing, "the usual", recommendations) lives skill-side; Bedrock calls go in a new adapters/bedrock_agent.py, never in core.

3. Persistence - swap SQLite for DynamoDB/RDS via a new adapters/dynamo_orders.py implementing the same function signatures as orders.py, selected by BARRIOS_ORDER_BACKEND.

4. Secrets - Secrets Manager / Lambda env vars, never hardcoded.

See adapters/aws_lambda.py for the full boundary notes.