An Agent OS That Lets Your Xiaomi Gear Think for Itself

What it does Anima is an open-source Agent OS that lives on your local network and treats smart hardware as collaborative agents rather than passive endpoints. It auto-discovers Xiaomi/Mi Home devices, maintains a layered long-term memory of your preferences and habits, and uses a LangGraph-based LLM Brain to plan actions. Each device type is wrapped in a Skill—a domain-knowledge package that includes comfort models, circadian lighting logic, and cross-device coordination rules rather than simple on/off toggles.

The interesting bit The project treats a Skill as a full lifecycle object: it can be created from natural language, registered, retrieved by the planner, executed, and refined through feedback. The memory system is deliberately tiered so the Brain only loads confirmed, evidence-backed details when a specific device Skill is invoked, avoiding the cost of dumping your entire history into every prompt.

Key highlights

• Auto-discovers Mi Home devices via local network scanning and Xiaomi Cloud QR login, skipping manual IP and token entry
• Brain runs on any OpenAI-compatible LLM and makes decisions inside explicit skill boundaries and safety rules
• Per-device Skills bundle prompts, domain knowledge, and action scripts for lights, humidifiers, ACs, air purifiers, and speakers
• Memory system extracts preferences from interaction history, tracks positive/negative evidence, and promotes only confirmed memories to active decision context
• Ships with a React dashboard, FastAPI backend, REST API, and CLI for realtime observation and debugging

Caveats

• Hardware support is currently limited to Xiaomi/MIOT devices; broader protocol support is listed as future work
• Ordinary device control still requires local network reachability and a device token; cloud login handles discovery and token acquisition, not full cloud remote control
• Xiaomi speaker cloud playback support is only partial

Verdict A promising experiment for Xiaomi-centric smart homes that want context-aware, learning automation instead of static if-this-then-that rules. If your ecosystem is outside Mi Home, there is not much to control yet.