Home Energy Sentinel

What it does

Home Energy Sentinel lets a conversational assistant prepare a safer time-of-use energy plan while keeping the final action under explicit human control.

The problem

Home-energy decisions combine changing tariffs, device deadlines, and a power limit. An assistant can find a useful plan quickly, but silently applying a plan is unsafe when the tariff or household state has changed.

The experience

The user asks what changed and what the assistant would do tonight. The agent reads the current synthetic state through MCP, identifies a cheaper overnight window, and proposes a sequential schedule for the water heater and EV. The active plan remains version 1 and the interface says HUMAN APPROVAL REQUIRED.

The person can approve or reject the proposal. Approval is a separate human operation, moves the plan from version 1 to version 2, and records a before/after audit event. Repeating the same approval is idempotent and does not create a duplicate action.

The demo then changes the synthetic tariff. The old proposal carries the previous tariff fingerprint, so approval returns STALE_PROPOSAL; the active plan remains unchanged. After rereading the state, the agent creates a fresh proposal against the new tariff version.

Why a conversational Alexa+ experience

Energy questions are natural spoken requests: “What changed?” and “Prepare the safest plan, but do not apply it yet.” Conversation is useful for discovery and explanation, while the visual surface makes the proposal, approval boundary, and provenance easy to inspect. The experience is deliberately multimodal: voice-like prompts start the workflow, and the screen exposes the state a person must review.

Architecture

The local project exposes three typed MCP tools over Streamable HTTP using MCP 2025-11-25:

• get_energy_state — read-only synthetic state;
• propose_energy_plan — creates a pending proposal and cannot mutate the active plan;
• approve_energy_plan — applies only from the human channel after checking tariff, consumption, and active-plan fingerprints.

The web simulation uses the same MCP server and a separate human-role client. The agent phase ends before approval. The policy harness is deterministic; no LLM, Alexa+ account, smart-home device, meter, or cloud service is involved.

Evidence and limits

The repository includes source, tests, the one-screen UI, transcripts, and reproducible evidence. npm test passes all 12 tests, including proposal non-mutation, explicit human approval/rejection, stale invalidation, and idempotency.

This is a synthetic, local demonstration. It does not claim a live Alexa+ integration, real energy savings, real devices, customers, or production authentication. The simulated path is used because Alexa+ Toolkit access was not available in this environment.

Work created for this challenge

The current Home Energy Sentinel project and its local UI/conversational workflow were created or significantly updated during the hackathon window. No prior Expediente Vivo, ClaimTrace, Decision Sentinel, or other project code is included in this isolated project.

Built With

• alexa+
• energy
• human-in-the-loop
• javascript
• mcp
• node.js
• streamable-http