SmartChain

Multi-provider LLM conversation agent for Home Assistant

Overview

SmartChain is a Home Assistant custom integration that provides an intelligent voice/conversation assistant powered by multiple LLM providers through LangChain. It also ships a sidebar panel for camera image analysis.

Supported providers:

• GigaChat (Sber) — Russian-focused LLM with vision support
• YandexGPT — Yandex Cloud LLM
• OpenAI — GPT-4.1, GPT-4o, o3, o4-mini
• Ollama — local models (Llama, Qwen, Gemma, T-Pro, DeepSeek, Home-3B)
• DeepSeek — cheapest cloud provider (V3, R1)
• Anthropic — Claude (Sonnet, Haiku, Opus)

Key Features

Conversation

• 6 LLM providers — cloud and local, switch without losing configuration
• Multiple agents — different models and prompts per provider (sub-entries)
• Streaming responses — real-time token-by-token output
• Device control — Assist API (tool calling): lights, switches, locks, climate
• Multi-agent orchestration (v4.4.0+) — ask_agents parallel fan-out across up to 5 siblings, critique_response second-opinion review, ask_agent single delegation
• Custom tools from YAML (v4.1.0+) — declarative LLM-callable tools with four action types (service, template, rest, script); per-subentry allowed_tools filter
• MCP client (v4.2.0+) — connect to remote MCP servers (stdio / sse / http) — filesystem, GitHub, brave-search, etc.; per-server auto-reconnect
• Long-term memory / RAG (v4.3.0+) — Chroma vector store; search_memory LLM tool; conversation + (opt-in) HA logbook ingest; pluggable embeddings (Ollama / OpenAI / GigaChat / Yandex)
• State history — get_state_history tool for past device states
• Vision — camera image analysis via multimodal models
• Skill system — loadable YAML files with additional knowledge
• Prompt caching — token savings on repeated requests
• Chat history — multi-turn conversations with context
• Jinja2 templates — customizable system prompt with device context

Services

• smartchain.ask — send a message to LLM from automations (Telegram, Slack, etc.)
• smartchain.analyze_image — camera snapshot → multimodal LLM → response
• smartchain.reload_tools (v4.1.0+) — re-read tools.yaml, restart MCP connections, rebuild memory subsystem atomically
• smartchain.clear_memory (v4.3.0+) — delete stored memories filtered by kind and/or agent_id
• AI Task entity — structured data generation for automations

SmartChain AI Panel

• Sidebar panel with a camera analysis tab
• Pick any HA camera, ask the LLM a natural-language question about the snapshot
• Result is mirrored to the smartchain_image_analyzed event and sensor.smartchain_last_analysis

Note: The YAML automation/script/scene/blueprint generation feature was removed in v4.0.0. See CHANGELOG.md for migration details.

Installation

Requirements

• Home Assistant 2024.12.0+
• HACS installed

Install via HACS

1. Add this repository as a custom HACS repository: https://github.com/dzerik/ha-smartchain
2. Search for "SmartChain" in HACS
3. Install and restart Home Assistant

Quick Start

1. Add Integration

Settings > Devices & Services > Add Integration > SmartChain

2. Select Provider and Enter Credentials

Provider	What you need
GigaChat	Auth credentials from developers.sber.ru
YandexGPT	API Key + Folder ID from Yandex Cloud
OpenAI	API key from platform.openai.com
Ollama	Base URL (default: http://localhost:11434)
DeepSeek	API key from platform.deepseek.com
Anthropic	API key from console.anthropic.com

3. Configure Options

• Model — select from list or type custom model name
• Assist API — enable device control via LLM tool calling
• System Prompt — customize the assistant's behavior
• State History Tool — enable past event analysis

4. Activate Assistant

Settings > Voice Assistants > Add — select your SmartChain entity as the conversation agent.

5. Open SmartChain AI Panel

Click SmartChain AI in the Home Assistant sidebar to open the camera analysis panel.

Documentation

Full user guide with all features and running examples:

• English: docs/USAGE.md
• Русский: docs/USAGE-ru.md

Covers: providers and credentials · subentry options · all services with examples · built-in conversation tools (Assist API, history, delegate, multi-agent, search_memory) · custom tools from YAML (service / template / rest / script) · MCP client (stdio / SSE / HTTP) · long-term memory (Chroma + 4 embedding providers) · AI Task entity · sidebar panel · skills system · troubleshooting.

What's new

Version	Highlights
v4.4.0	Multi-agent orchestration — ask_agents parallel fan-out + critique_response second-opinion review
v4.3.0	Long-term memory / RAG — Chroma vector store, search_memory tool, conversation + logbook ingest
v4.2.0	MCP client — connect to remote MCP servers via stdio / SSE / HTTP with auto-reconnect
v4.1.0	Custom tools from YAML — declarative LLM tools (service / template / rest / script)
v4.0.2	Security fixes, proper Last Analysis SensorEntity, correct integration_type
v4.0.0	Major refactor — focused on conversation, AI Task and vision

Full release notes: CHANGELOG.md.

License

MIT