Steward - an Alexa+ Agent Skill

Steward is a personal-operations Agent Skill for the Alexa+ track of the Build, Ship, Shape: Amazon Developer Hackathon. It runs errands across your calendar, reminders, notes, and remembered preferences through a self-hosted MCP server (Streamable HTTP, MCP spec 2025-11-25), orchestrated by a Strands SDK agent. A small web app provides the simulated Alexa+ experience allowed by the hackathon rules.

The agent also enters the AWS Builder mini-challenge (Strands SDK agent framework plus a genuine Amazon Bedrock model path with documented setup) and the Open Source mini-challenge (this repo: new, MIT-licensed, built during the hackathon window).

How it works

You type an errand
      v
Web UI (simulated Alexa+ experience, FastAPI)
      v
Strands SDK agent  (AWS Builder: Strands + Bedrock integration)
      |  MCP client (Streamable HTTP)
      v
Steward MCP server (13 tools, SQLite store, zero cloud dependencies)

Example: "Plan my Saturday: dentist at 10am, buy groceries, call mom" makes the agent call clock_now, calendar_add, reminder_add, and note_save, then summarize the plan. Preferences persist across sessions via prefs_set / prefs_get (the agent checks them before assuming anything).

Related MCP server: konbu

Setup

Requires Python 3.10+.

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

Configure the model

The agent needs a model. Pick one:

Option A: Amazon Bedrock (the AWS path). Configure AWS credentials the usual way (env vars, ~/.aws/config, or IAM role), then:

export STEWARD_BEDROCK_MODEL_ID=amazon.nova-micro-v1:0   # or any Bedrock id
export STEWARD_AWS_REGION=us-east-1

Option B: any OpenAI-compatible endpoint (local server, gateway, etc.):

export STEWARD_MODEL_BASE_URL=http://127.0.0.1:11434/v1
export STEWARD_MODEL_ID=your-model-id
export STEWARD_MODEL_API_KEY=your-key   # or "not-needed"

With AWS credentials present and no STEWARD_MODEL_BASE_URL, Bedrock is selected automatically. Set STEWARD_MODEL_PROVIDER=bedrock or STEWARD_MODEL_PROVIDER=openai-compatible to force one. With nothing configured, the agent raises a clear error instead of guessing.

Run it

Terminal 1, the MCP server (Streamable HTTP on 127.0.0.1:8899):

python -m steward.mcp_server

Terminal 2, the web experience:

python -m steward.web

Open http://127.0.0.1:8898/ and give Steward an errand, for example:

• "Plan my Saturday: dentist at 10am, buy groceries, and remind me to call mom"

• "Remember that I prefer morning appointments"

• "What is on my calendar tomorrow?"

The UI shows which MCP tools ran for each reply.

Talk to the MCP server directly

Any MCP client works. With the Python MCP SDK:

import asyncio
from mcp.client.session import ClientSession
from mcp.client.streamable_http import streamable_http_client

async def main():
    async with streamable_http_client("http://127.0.0.1:8899/mcp") as (r, w):
        async with ClientSession(r, w) as s:
            await s.initialize()
            tools = await s.list_tools()
            print([t.name for t in tools.tools])

asyncio.run(main())

MCP tools (13)

Tests

python -m pytest tests/ -q

21 tests: SQLite store CRUD, the MCP server over real Streamable HTTP (protocol version asserted >= 2025-11-25), provider selection, the Strands agent loop against the real server with a scripted model double, and the web API.

Project layout

steward/
  store.py       SQLite storage (events, reminders, notes, prefs)
  mcp_server.py  MCPServer "steward" v1.0.0, 13 tools, Streamable HTTP
  models.py      build_model(): Bedrock or OpenAI-compatible (Strands)
  agent.py       StewardAgent: Strands agent + MCP client + session memory
  web.py         FastAPI app: the simulated Alexa+ experience
  netenv.py      proxy-env sanitizer for localhost MCP clients
webui/index.html Single-page chat UI (no build step)
tests/           pytest suite (21 tests)

Environment variables

License

MIT. See LICENSE.