What I'm trying to build here: a home-energy manager that's versatile, precise, and made to run for years — and the tooling to set systems like it up efficiently with AI. Open source, work in progress.
The approach across all of it is cautious by design: observe before acting, verify before trusting, small steps to earn confidence. I also try not to reinvent things — everything here is something I needed, couldn't find anywhere, and so ended up building myself. On that mission I found (to my knowledge) new ways to use AI agents to build code efficiently and trustworthily (I learn something every day).
- hemm / ha-hemm — reads device manifests, constraints, and price/solar forecasts and plans 24 h of power for Home Assistant. It observes and plans first; taking control is opt-in. Soft-beta: it installs and runs, but it's still moving.
- hactl — a CLI for driving Home Assistant from LLM agents, token-efficient by design. The most finished thing here.
- hactl-companion — adds create/update/delete of HA entities on top of hactl.
- Cross-model agentic development — one model (Claude Opus) plans and reviews while another (OpenAI Codex) writes the code, so every change is read by a model other than the one that wrote it. The reason I think it helps: a model waves through its own output but picks apart another's, so review across two models catches more than self-review. → the full writeup
- Testing the EMS against a living Home Assistant — the integration tests run against a real, ephemeral HA instance in Docker via hactl, not mocks, so the plans are checked against the system that actually runs them.
Built by Jan Kipping, embedded and automotive software engineer.
