Home Memory & Mechanic

Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for Home Memory & Mechanic

Inspiration

I'm a Florida home inspector, in houses almost every day for insurers and homeowners. My 4-point and full inspection reports get read once, at closing or renewal, then forgotten. Yet they're the best record of a home's systems. So I put the inspector in the house.

What it does

Home Memory & Mechanic is an MCP server for Alexa+. It learns a house from photos of its data plates, remembers it, and helps keep it insurable and running.

• Scan: scan_component reads a data-plate photo for brand, model, serial, capacity and refrigerant, each with a confidence score. It decodes the manufacture date and asks the year when it isn't sure. confirm_component saves it, and the owner's date always wins.
• Home profile: every system's age against its expected life, plus consumables and what's due.
• Insurability check: roof, electrical panel, wiring, plumbing, water heater and HVAC are each rated Ready, At risk, Likely problem, or Need info, with a fix list. Florida's 4-point is named only for Florida homes. This is not an inspection.
• Triage: guided checks with photo steps for an AC not cooling, a dead outlet, or a chirping smoke alarm. Safety stops end the repair path and write a ticket.
• Exact-fit consumables: filters and alarms, with a cart link only after a yes and never a checkout.
• Maintenance calendar: seasons follow the home's hemisphere. Recall check against the US CPSC list.
• Service tickets: drafted and sent only after an explicit yes.
• Six MCP Apps cards: spoken answers stay short, and the cards carry the detail.

How we built it

• Server: Python 3.12, uv, and the MCP Python SDK (mcp==2.3.0). Protocol 2025-11-25 over Streamable HTTP, with bearer-key auth.
• Surface: 13 tools, 4 resources (home://profile, home://components/{id}, home://reports/insurability/latest, home://history), 3 prompts, and 6 MCP Apps cards rendered from ui:// resources. Tests assert all of these.
• Photos: uploaded first, so the tools receive only an image key.
• Simulator: our own web app styled like an Echo Show, with browser speech, a simulator backend and a Node front end. It is not Amazon's partner Web Simulator.
• Assistant: a Claude tool-use loop when BEDROCK_MODEL_ID is set (Anthropic SDK on Bedrock). Otherwise an offline router and a scripted plate reader (DEMO_MODE).
• Storage: locally, a JSON home store and a photo folder. AWS-ready: DynamoHomeStore and S3PhotoStore with presigned uploads, tested under moto, and a CDK stack that synthesizes the tables, bucket, KMS key and bearer secret. It isn't deployed.
• Rules: shared with Vacant Home Watch through home-conditions-engine and the open-source home-conditions library.
• Tests: 80 tests passing at the last review.

Challenges we ran into

• Partner-only tools. The Alexa+ Category SDK, MCP Toolkit, CLI and Web Simulator read as available, then turn out to be partner-only (F-001). We built our own simulator.
• Import exits the process. Importing mcp.cli calls sys.exit(1) when typer is missing (F-002).
• Tools silently dropped. Tools registered after MCPServer is built are dropped with no warning (F-003).
• Elicitation fails on stateless transport. It raises NoBackChannelError on the stateless transport even though the client advertises elicitation (F-004). We added a Confirm-tap or second-call fallback.
• Bare 421. A non-localhost Host gets a bare HTTP 421 (F-005).
• No container for AgentCore. AgentCore Runtime needs a container image, and Docker isn't installed. The hosting choice is still open.

Accomplishments that we're proud of

• A full MCP server on the newest spec, with cards, elicitation-based consent, and memory across sessions. Restart the server and it still knows the house.
• Hard safety lines: nothing opens a panel, touches gas or refrigerant, or mixes water and electricity. Nothing is bought or sent without a yes.
• Shared rules: the same inspector rules engine runs in both of my entries.

What we learned

• On a voice surface, short spoken answers plus cards beat long answers.
• Consent has to work on every transport: elicitation where the client supports it, otherwise a Confirm tap or a spoken yes.
• When a date is uncertain, asking beats guessing.

What's next

All planned in the repo:

• a deployed MCP URL on AWS
• Strands as the assistant, on AgentCore Runtime with AgentCore Memory
• real Amazon product IDs, instead of spec searches
• Polly and Transcribe voice
• EventBridge maintenance nudges

Built With

• amazon-bedrock
• amazon-dynamodb
• anthropic-sdk
• aws-cdk
• aws-kms
• aws-secrets-manager
• cpsc-recalls-api
• mcp
• mcp-apps
• model-context-protocol
• moto
• node.js
• pytest
• python
• streamable-http