Smart Home MCP Lab

Governed agent access to real devices — remote MCP over AWS (Cognito OAuth, Lambda, IoT Core) plus a local agent loop with persistent memory and a Slack front-end.

Control Paths

Related MCP server: MCP MQTT Light Controller

Features

MCP Paths

• Turn lights on/off, set brightness, get bulb status (on/off, brightness, color temp)

• Real bulb control via tapo library over local network

• Mock fallback — automatically uses a mock when credentials are missing or the bulb is unreachable

• Persistent state — bulb state survives server restarts

• DynamoDB state logging — fire-and-forget, never blocks MCP tools

• AWS IoT Core integration — MQTT bridge with Device Shadow for state sync

• AgentCore Gateway — remote MCP server with Cognito OAuth for Claude web app

• Multi-device support — single bridge manages multiple devices via DeviceRegistry

• Device-agnostic architecture — BaseDevice interface makes adding new device types straightforward

Local Agent

• Interactive CLI and Slack bot — two front-ends sharing the same agent loop, memory, and skills

• Slack Socket Mode — per-thread sessions, @mention in channels, direct messages; idle sessions auto-evict after 30 min with memory flush

• Multi-provider LLM support — swap between Anthropic (Claude) and Google Gemini by passing --model; provider auto-detected from the model name prefix

• Persistent memory, no cloud dependency — Markdown files + SQLite index, runs entirely on-device

• Hybrid memory search — BM25 (FTS5) + vector embeddings (ollama) merged via Reciprocal Rank Fusion

• Pluggable skills — drop a SKILL.md into skills/<name>/ — no Python scripts, no boilerplate, zero changes to the loop

• Progressive skill disclosure — only a skill index (name + description) is in the system prompt at startup; full docs load on first use via describe_skill, then stay injected for the session

• Direct command execution — run_command tool lets the model run any shell command and parse JSON output; skills are pure documentation that describe which commands to run

• Color temperature control — set_color_temp (2500–6500 K) via the agent skill

• Heartbeat scheduler — HeartbeatScheduler fires time-based automations from SCHEDULE.md; tasks store a shell command (cmd) and run as subprocesses; Claude manages tasks via schedule_task (add/remove/list), changes persist across restarts

Project Structure

src/smarthome/
├── devices/              # Shared device layer (all paths use this)
│   ├── base.py           # BaseDevice ABC: execute(), apply_desired_state(), get_shadow_state()
│   ├── device_registry.py # Manages multiple devices by ID
│   ├── tapo_bulb.py      # TapoBulb (real hardware) + MockTapoBulb (testing/fallback)
│   └── bulb_cli.py       # CLI entry point for agent skills (outputs JSON, reads TAPO_MOCK)
├── aws_mcp/              # AWS path: Local MCP server + Lambda + IoT bridge
│   ├── bridge/           # IoT Core MQTT bridge (config, iot_bridge, shadow_manager)
│   ├── cloud/            # Lambda-side IoT helpers (iot_commands)
│   ├── logging/          # DynamoDB state logger
│   ├── mcp_servers/      # FastMCP server (light_server.py)
│   └── lambda_handler.py # AgentCore Gateway Lambda entry point
└── agent/                # Local-first agent loop (CLI + Slack)
    ├── __main__.py       # Entry: `python -m smarthome.agent [--mock|--slack]`
    ├── config.py         # AgentConfig: paths, model, mock flag
    ├── loop.py           # AgentLoop: provider-agnostic tool-use loop + 6 built-in tools
    ├── scheduler.py      # HeartbeatScheduler: fires SCHEDULE.md cmd tasks as subprocesses
    ├── skill_loader.py   # Discovers skills/*/SKILL.md, builds system prompt index
    ├── slack_adapter.py  # Slack channel adapter (Socket Mode, per-thread sessions)
    ├── providers/        # LLM provider abstraction
    │   ├── types.py      # NeutralTool, ToolCall, ToolResult, ProviderResponse
    │   ├── base.py       # Abstract LLMProvider
    │   ├── anthropic.py  # AnthropicProvider
    │   └── google.py     # GoogleProvider (Gemini)
    ├── memory/
    │   ├── manager.py    # MemoryManager: search, write, sync, session context
    │   ├── schema.py     # SQLite schema: files, chunks, FTS5, vec, device_events
    │   ├── chunker.py    # Markdown → overlapping chunks (~400 tokens)
    │   └── embedder.py   # OllamaEmbedder: async HTTP → ollama /api/embed
    └── skills/
        ├── light-control/
        │   └── SKILL.md  # Skill docs: commands to run via bulb_cli.py
        └── trading-journal/
            └── SKILL.md  # Skill docs: analytics pipeline commands

scripts/aws/              # AWS provisioning and operation scripts
docs/                     # Setup guides and architecture notes
tests/                    # Unit tests (200 tests, all passing)

How It Works

Local MCP (Claude Desktop)

Claude Desktop runs the FastMCP server as a local subprocess:

Claude Desktop → FastMCP server (subprocess) → TapoBulb / MockTapoBulb

Tools exposed: turn_on, turn_off, set_brightness(level), get_status

On first tool call the server tries to connect to a real bulb using credentials from ~/.smarthome/.env. Falls back to mock automatically if credentials are missing or the bulb is unreachable.

Remote MCP (Claude Web App)

Claude Web App
  → AgentCore Gateway (Cognito JWT auth)
  → Lambda (smarthome-gateway-handler)
  → IoT Core MQTT
  → IoT Bridge (local network)
  → TapoBulb

See docs/claude-web-oauth.md for the OAuth flow details and docs/mcp-setup.md for full provisioning steps.

Local Agent

An OpenClaw-inspired agent loop that runs entirely locally — no cloud dependency. Two front-ends share the same AgentLoop, memory, and skills:

CLI input   →  AgentLoop  →  LLMProvider (Anthropic or Google Gemini)
Slack input ↗     ├── memory/  ~/.smarthome/memory/ — MEMORY.md, USER.md, SOUL.md, daily logs
                  └── skills/  light-control, trading-journal — SKILL.md only, no scripts

Front-ends:

• CLI — interactive REPL, one session per process

• Slack (--slack) — Socket Mode bot; one session per (channel, thread), responds to @mention in channels and all messages in DMs; idle sessions auto-evict after 30 min with memory flush

6 built-in tools the model can call:

1. run_command(cmd) — run any shell command; parses JSON stdout; lifts Slack blocks

2. describe_skill(skill_name) — loads full skill docs once; injected into system prompt for the session

3. memory_search(query) — hybrid BM25 + vector search (Reciprocal Rank Fusion)

4. memory_write(path, content, mode) — persists to Markdown files

5. schedule_task(action, name, hour, minute, cmd) — add/remove/list scheduled automations in SCHEDULE.md

Provider abstraction: tools and conversation history are stored in a neutral format. AnthropicProvider and GoogleProvider each serialize to their own wire format on every call. Provider-specific data that must survive conversation round-trips (e.g. Gemini's thought signatures when thinking is enabled) is carried opaquely in ProviderResponse.assistant_replay.

Memory is stored in ~/.smarthome/memory/ as Markdown files (MEMORY.md, USER.md, SOUL.md, daily logs), indexed in SQLite with FTS5 and optional sqlite-vec embeddings. Embeddings use ollama; BM25-only fallback if unavailable.

Adding a skill: drop a folder under skills/, write SKILL.md with frontmatter (name, description) and shell command examples. No Python scripts, no boilerplate. Zero changes to loop.py.

Device Layer

All devices implement BaseDevice:

• execute(action, parameters) — dispatch any action (turn_on, set_brightness, …)

• apply_desired_state(desired) — apply state from IoT Shadow delta

• get_shadow_state() — report current state to shadow

TapoBulb connects to real hardware. MockTapoBulb simulates a bulb in memory, optionally persisting state to ~/.smarthome/tapo_bulb_state.json.

Setup

See docs/mcp-setup.md for full step-by-step instructions covering both paths.

Quick start — Local MCP

1. Create ~/.smarthome/.env with bulb credentials (or skip — mock mode works without it):

   TAPO_USERNAME=your_tapo_email
   TAPO_PASSWORD=your_tapo_password
   TAPO_IP_ADDRESS=192.168.x.x

2. Add to Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json):

   {
     "mcpServers": {
       "smarthome": {
         "command": "uv",
         "args": ["run", "--directory", "/path/to/smarthome",
                  "fastmcp", "run",
                  "src/smarthome/aws_mcp/mcp_servers/light_server.py"]
       }
     }
   }

3. Restart Claude Desktop.

Quick start — Local Agent

1. Add your API key(s) to ~/.smarthome/.env:

   mkdir -p ~/.smarthome
   
   # Anthropic (default model: claude-sonnet-4-6)
   echo 'ANTHROPIC_API_KEY=sk-ant-...' >> ~/.smarthome/.env
   
   # Google Gemini (optional — use with --model gemini-*)
   echo 'GEMINI_API_KEY=AIza...' >> ~/.smarthome/.env

2. Seed memory files (optional but recommended):

   mkdir -p ~/.smarthome/memory
   echo "# Memory" > ~/.smarthome/memory/MEMORY.md
   echo "# User Preferences" > ~/.smarthome/memory/USER.md

3. Run with mock bulb (no hardware needed):

   uv run python -m smarthome.agent --mock                               # Claude (default)
   uv run python -m smarthome.agent --mock --model gemini-3.1-flash-lite # Google Gemini

4. Run with real bulb — add TAPO_USERNAME, TAPO_PASSWORD, TAPO_IP_ADDRESS to ~/.smarthome/.env, then:

   uv run python -m smarthome.agent
   uv run python -m smarthome.agent --model gemini-3.1-flash-lite

5. Run as a Slack bot — add SLACK_BOT_TOKEN, SLACK_APP_TOKEN, SLACK_SIGNING_SECRET to ~/.smarthome/.env, then:

   uv run python -m smarthome.agent --slack --mock   # mock bulb
   uv run python -m smarthome.agent --slack          # real bulb

Quick start — Remote MCP (AWS)

Run provisioning scripts in order (requires AWS profile self):

AWS_PROFILE=self uv run python scripts/aws/create_bridge_thing.py
AWS_PROFILE=self uv run python scripts/aws/create_cognito.py
uv run python scripts/aws/package_lambda.py
AWS_PROFILE=self uv run python scripts/aws/create_lambda.py
AWS_PROFILE=self uv run python scripts/aws/create_agentcore_gateway.py

# Start the local bridge (keep running on-premises)
uv run python scripts/aws/run_bridge.py

# Test end-to-end
AWS_PROFILE=self uv run python scripts/aws/test_gateway.py

Testing

# Unit tests
uv run pytest tests/ -v

# Interactive MCP dev UI (localhost:6274)
uv run fastmcp dev src/smarthome/aws_mcp/mcp_servers/light_server.py

Key Dependencies

Dependencies are split so the local agent can be installed without the AWS/MCP stack.

Core (local agent):

aws-mcp extra (MCP paths only):

Dev:

Install for each environment:

uv sync --no-dev             # Raspberry Pi — local agent only
uv sync --extra aws-mcp      # Dev machine — everything

What's Next

• Local MCP via Claude Desktop

• Remote MCP via AgentCore Gateway + Cognito OAuth

• Multi-device support via DeviceRegistry

• Local agent loop with Markdown memory

• Bulb control as an agent skill

• Color temperature control

• Heartbeat scheduler with SCHEDULE.md and schedule_task tool

• Multi-provider LLM support (Anthropic + Google Gemini)

• Interoperable skills — SKILL.md-only, executable via run_command (no Python scripts required)

• Additional device types (smart plugs, sensors)

• Device auto-discovery on local network

License

MIT