Two months of nights and weekends. Ready to share.

Claw Assistant hooks into your existing OpenAI / Claude / Gemini / OpenRouter conversation pipeline. Doesn't replace anything. Same setup

Privacy and safety first:

Your AI only sees entities you explicitly expose. All config changes are staged — AI proposes, you confirm, then it applies. Nothing writes to your system silently.

What it can do:

Dashboard from a screenshot upload a UI mockup or describe what you want. It queries your entities, picks cards, builds the view.

Automation with self-verification creates the automation, then pulls the trace to confirm it actually fired. Tells you exactly what failed and why.

Frontend control navigates HA pages, reads the DOM, clicks elements, injects JS. Like a real user verifying what rendered.

Web search mid-conversation searches, reads full pages, brings info back. No window switching.

Safe file ops staging → confirm → apply. Reversible.

Cross-session memory remembers your areas, device names, preferences. No re-explaining your setup every time.

Skill system good patterns get saved as reusable skills. Gets better with your specific setup over time.

HACS management install/update/uninstall from the conversation. Frontend resources auto-reload after install.

Multi-agent agents consult each other mid-turn. Subtasks delegated, results merged.

vs. MCP tools (ha-mcp- etc.):

MCP runs outside HA, exposes ~70 atomic API wrappers to external clients that have zero context about your instance. High token cost, many round-trips.

Claw runs inside HA with full registry access, area structure, trace history, live states. One instruction = multi-step coordinated workflow. Also knows voice vs. text — no markdown walls in your TTS.

Install: HACS custom repo or manual. Any tool-calling LLM.

https://github.com/ha-china/ha_claw/

[留言]

为什么值得关注

能改变理解方式，而不只是重复常识；符合当前抓取需求；它提供了新的理解或解释，而不只是表面观点

来源：reddit，领域：tech，保留分：0.57