Routing
Right model for the job.
Simple commands skip the LLM. Complex requests get a frontier model. Conversation uses a tuned middle tier. Latency where it matters; reasoning where it counts.
The Agent
Your house has an opinion now.
The Butler is the agent inside Elysium. It reasons about your home, picks the right action, and explains itself when you ask. Built on LangGraph; runs in the cloud by default, on your own GPU when you'd rather.
What it does.
The Butler listens for natural-language requests, picks the right tool, and executes — turning “dim the lights and play something quiet” into two distinct actions. It thinks in steps, can explain its plan, and verifies each move before reporting back.
State persists across turns. Ask a follow-up without restating context. Conversations are checkpointed so the agent recovers from crashes mid-step instead of losing the thread.
How it works.
Reasoning
Built on LangGraph.
Stateful graph with checkpointed turns, crash recovery, and inspectable steps. The Butler's reasoning is a thing you can read, not a black box.
Sovereignty
Cloud or local.
Anthropic-hosted by default so it works out of the box. Point it at a Qwen3 14B AWQ instance on your own GPU when you'd rather your home stays on your network.
When it matters
Sees the moment. Calls for help.
The Butler doesn't wait to be asked. When something changes in the home, it acts — quietly, decisively, and stays present until the humans arrive.
Run it your way.
Elysium ships with Anthropic-hosted models so it works out of the box. Want full local control? Point the agent at vLLM on a consumer GPU — an NVIDIA 5090 fits the whole stack (Qwen3 14B AWQ + Whisper + Piper TTS + YOLO vision) with headroom — and your home stays on your network.
Want early access?
We're shipping Elysium to a handful of homes first. Drop us a line if you'd like to be on the list.
hello@elysium-labs.ai