A cognitive substrate for Alexa+: an MCP server that gives neurodivergent people persistent memory, a sense of time, a translator for ambiguous messages, task breakdown, a low-overwhelm daily plan, and guardrails against rumination and hyperfocus. Because hackathon entrants can't deploy to real Alexa+, it ships with an Echo Show–style Alexa+ simulator: a real MCP client with real account linking. A Telegram companion bot for forwarded messages and opt-in nudges is designed (spec §5.7) but not built yet.
See the spec for design, tools, and AWS architecture, and AWS setup for account, Bedrock, Cognito, and agent-toolkit configuration.
Built for the Amazon Developer Hackathon 2026 — Alexa+ track.
▶ Demo video (2:52): https://youtu.be/tjRjl7ehqKg
Implements a Model Context Protocol server over the Streamable HTTP transport, per the track requirements.
- Python 3.13+
- uv
git clone https://github.com/navinrc/NeuroCare.git && cd NeuroCare
uv sync
cp .env.example .env # optional: HOST / PORT
uv run alexa-mcp-seed --reset # load synthetic demo data (profile_seed.yaml) into .data/
uv run alexa-mcp-server # http://127.0.0.1:8000/mcpWithout auth env vars every request is served as DEMO_USER_ID (default demo). To require a
bearer token locally:
ALLOW_STATIC_TOKEN=true AUTH_TOKEN=devtoken uv run alexa-mcp-serverInspect with the MCP Inspector:
npx @modelcontextprotocol/inspector
# Transport: Streamable HTTP, URL: http://127.0.0.1:8000/mcp
# With a static token: add header Authorization: Bearer devtoken
# Try: what_was_i_doing, recall("my boss"), open_loops| Variable | Default | Notes |
|---|---|---|
HOST, PORT |
127.0.0.1, 8000 |
|
STORE, SQLITE_PATH |
sqlite, .data/neurocare.sqlite3 |
|
ALLOW_STATIC_TOKEN, AUTH_TOKEN |
off | Local dev / Inspector bearer token |
DEMO_USER_ID |
demo |
User served when auth is off |
LLM_PROVIDER |
off |
bedrock turns on the LLM refine and the embedding tier; off = deterministic baselines only |
BEDROCK_MODEL_ID |
us.anthropic.claude-haiku-4-5-20251001-v1:0 |
Converse model / inference profile |
BEDROCK_EMBED_MODEL_ID |
amazon.titan-embed-text-v2:0 |
|
AWS_REGION |
us-east-1 |
Credentials come from the usual AWS chain (AWS_PROFILE, …) |
TABLE_NAME |
cognitive-substrate |
DynamoDB table when STORE=dynamo |
COGNITO_USER_POOL_ID, COGNITO_CLIENT_ID, COGNITO_SCOPE |
unset | Cognito access tokens (Alexa+ account linking); sub is the user id |
PUBLIC_BASE_URL |
http://HOST:PORT |
Resource URL in the OAuth metadata; on Lambda it defaults to the Function URL |
Hackathon entrants get no Alexa+ developer tools, so the demo runs on an Echo Show–style simulator (spec §5.9): a real MCP client of the server with real Cognito account linking, Bedrock tool use, voice in/out (Chrome), and native reminders.
# local: server with a static token, simulator using it (Bedrock credentials needed for the model)
ALLOW_STATIC_TOKEN=true AUTH_TOKEN=devtoken uv run alexa-mcp-server
SIM_STATIC_TOKEN=devtoken AWS_PROFILE=<profile> uv run alexa-sim # http://127.0.0.1:8100Deployed: the SimulatorUrl stack output. Click Link NeuroCare, sign in as the demo user, then
hold the mic button (or space bar) and talk.
infra/ is the CDK app (DynamoDB, Lambda + Function URL, Cognito). Steps, account linking, and
the latency check: infra/README.md and
docs/verify/phase-6.md.
uv run alexa-mcp-inbox-token # prints a personal token once (a planned Telegram /token command would do this)
curl -X POST http://127.0.0.1:8000/inbox -H "Authorization: Bearer $TOKEN" \
-H 'Content-Type: application/json' -d '{"text": "Per my last email...", "sender": "Priya"}'
# then: decode() with no message, or triage_inbox()uv run pytest
BEDROCK_LIVE=1 AWS_PROFILE=neurocare-admin uv run pytest -m live -s # one real Bedrock call
scripts/prove.sh scripts/smoke/phase3.yaml # live HTTP proof (docs/verify/)
uv run ruff check .
uv run ruff format .src/alexa_mcp_server/
server.py # MCPServer instance, thin tool layer, entrypoint
config.py # env settings
auth.py # static dev token (Cognito JWT in phase 6)
store/ # Store interface + SQLite backend
pillars/memory.py # memory tools, profile, seed loader
profile.py # profile schema, defaults, validation
pillars/sense_of_time.py # now_context, work blocks, check_in
pillars/translator.py # decode, decode_result, triage_inbox, inbox items
pillars/inbox.py # personal inbox tokens for POST /inbox
llm/ # LLM + Embedder protocols, Bedrock clients, fakes for tests
refine.py # fast-first background refine jobs
semantic.py # embedding tier (recall, close_loop; rumination in phase 4)
matching.py # tiered recall: exact → alias → fuzzy → embedding
speech.py # short spoken-text helpers
seed.py # alexa-mcp-seed CLI
profile_seed.yaml # synthetic demo data
tests/ # pytest; in-process MCP client + pillar tests
MIT — see LICENSE.
