CareCircle: Alexa+ that remembers, so families don't have to worry

CareCircle is a self-hosted MCP server (spec 2025-11-25, Streamable HTTP) that gives Alexa+ a long-term memory and the judgment to look after an older adult who lives alone. Margaret, 78, just talks to Alexa the way she always has. CareCircle remembers her medicines, how she feels and when her pills run low, reminds her at the right time, and decides on its own when her daughter and son need to know.

Built for Build, Ship, Shape: Amazon Developer Hackathon: Alexa+ track (self-hosted MCP server and a simulated Alexa+ experience), plus the AWS Builder and Open Source mini challenges.

Why

About 1 in 4 adults over 65 lives alone, and missed medication is one of the most common reasons older adults end up in hospital. Families either call every day to ask "did you take your pills?" or find out too late. A voice assistant is already in the kitchen. It needs memory, timing and judgment.

Related MCP server: hearth-mcp

What it does

The judgment (risk engine, server/store.py)

After every dose and check-in, CareCircle decides by itself whether to escalate:

The same trigger is de-duplicated for 6 hours so families aren't spammed. Everything persists in SQLite, so the memory survives sessions, devices and restarts.

MCP surface

• 15 tools: whats_next, log_medication, daily_checkin, check_nudges, get_family_messages, get_refill_status, record_refill, prepare_doctor_visit, get_wellbeing_summary, alert_family, show_care_dashboard, get_medication_schedule, add_medication, add_appointment, setup_person

• MCP Apps extension (io.modelcontextprotocol/ui): show_care_dashboard → ui://carecircle/dashboard.html (text/html;profile=mcp-app). The simulator is a minimal MCP Apps host: it reads the resource and posts the tool result into a sandboxed iframe (ui/notifications/tool-result).

• Resource template: carecircle://people/{name}/profile

• Prompt: morning_routine

• Every tool returns speech (what Alexa says), card (what a screen shows) and structured data.

• DNS-rebinding protection stays on in production: only localhost and the deployed hostname are accepted.

Run it locally (2 minutes)

Requires Python 3.10+.

./run.sh

Then open http://localhost:8000. The live demo page shows three things side by side:

• Margaret's kitchen: an Echo Show–style Alexa+ simulator with a mic (Chrome), spoken replies, cards and proactive reminders

• Priya's phone: the CareCircle family app, updating live (also at /family)

• Under the hood: every MCP request, live, with arguments, latency and results

The demo clock buttons (8 AM / 1 PM / 8 PM) shift CareCircle's clock so reminders fire on cue. Reset demo restores a fresh week of history.

MCP endpoint: http://localhost:8000/mcp

npx @modelcontextprotocol/inspector     # connect to http://localhost:8000/mcp (Streamable HTTP)

from mcp import Client
async with Client("http://localhost:8000/mcp") as c:
    r = await c.call_tool("whats_next", {"person": "Margaret"})
    print(r.structured_content["speech"])

Deploy

Render (free, ~3 minutes): click the Deploy to Render button above, sign in with GitHub, and press Apply. render.yaml configures everything. You get https://carecircle-xxxx.onrender.com, with the MCP endpoint at /mcp. On the free plan the service sleeps after 15 minutes idle, and the first request takes about 30 s to wake it.

Docker (anywhere):

docker build -t carecircle . && docker run -p 8000:8000 -e PUBLIC_HOSTS=your.domain.com carecircle

PUBLIC_HOSTS (comma-separated) is the list of hostnames the MCP endpoint accepts besides localhost. Render's hostname is picked up automatically.

The Alexa+ agent's brain (Amazon Bedrock)

simulator/agent.py connects to CareCircle as a real MCP client and picks its model automatically:

1. Amazon Bedrock (Converse API with tool use): when AWS credentials are present. Model: CARECIRCLE_BEDROCK_MODEL (default us.anthropic.claude-haiku-4-5-20251001-v1:0).

2. Anthropic API: if ANTHROPIC_API_KEY is set.

3. Local planner: deterministic intent routing, so the demo runs with no keys.

Whichever brain runs, every action is an MCP tools/call. Force one with CARECIRCLE_BRAIN=bedrock|anthropic|local. Optional SMS/email to family through Amazon SNS: set CARECIRCLE_SNS_TOPIC_ARN. Without it, alerts go to data/outbox.jsonl.

Project layout

server/carecircle_server.py   MCP server: 15 tools, MCP App, resource, prompt, website + APIs (one process)
server/store.py               SQLite memory, risk engine, reminders, refills, messages, doctor brief
server/notify.py              family alert delivery (outbox / Amazon SNS)
server/ui/dashboard.html      family app, also served as the MCP App resource
simulator/agent.py            Alexa+ agent: MCP client + Bedrock / Anthropic / local brain
web/index.html                live demo: Echo Show simulator, family phone, MCP inspector
scripts/seed_demo.py          demo persona + a week of history
render.yaml, Dockerfile       deployment

Limits (honest)

• CareCircle is a reminder and family-communication tool, not a medical device. It doesn't diagnose, and a red-flag alert tells family; it doesn't call emergency services.

• The Alexa+ device is simulated (the hackathon's official simulation path). The MCP server is real and works with any MCP 2025-11-25 client.

• The demo is a single household with no sign-in, and the public demo is shared by everyone who opens it. Next step: OAuth 2.1 per household (supported by the MCP SDK).

License

MIT, see LICENSE.