I have two hubitat hubs (hub mesh) running alongside Home Assistant. I have been fine tuning my integration of AI (ChatGTP in this case) with Home assistant for Direct Control / reporting / writing automations WITH strict guidelines for it. Along the way, I discovered the solution I am using for HA also works with Hubitat WITHOUT any AI integration with Hubitat. The key is running it on my local PC using Opencode. I asked my AI through opencode if it could read or write any rule machine rules - and it said it could. How? I asked. It told me - it uses Python playwright controlling headless microsoft edge against the local hubitat IP address. So I gave it some tests. Wow. Now - I DO use guardrails on this stuff - opencode has a long agents.md document I wrote which governs what it can and can't do, and my operational preferences. But I am amazed now by what this all can do. Anyone else this far along? I don't mean asking AI to help YOU write automations - but having AI do the actual work.
I have done something similar with Hermes Agent using OpenAI with GPT5.6-luna. I just pointed it to the URL and it asked it to perform tasks. Hermes agent has the ability to use a local web browser to interact with websites. I also set it up to access the official Hubitat AI MCP Server. In my case i was asking it to interate on a driver and it would do so as needed, test it, review logs, and then make changes if needed. It was impressive what it could do. Though that was drivers i am sure it would of figured out if i asked it to create RM rules.
Haven’t tried it myself yet, but I believe this MCP server integration can write rules:
Thanks for the shout-out!
So yes it is possible to do all that on your local network over http using curl/bash/etc, but rule machine is extraordinarily complex and AI is also very prone to making many overly confident mistakes, so I'd be very careful with what you do if going that route.
Going thru an mcp server like my Hubitat Local MCP server gives a lot more control, safeguards as well as direct guidance for the agents, which makes it a lot safer imo, and also advertises the capabilities. It also makes it less likely to make mistakes because the instructions are right there in front of it, but to be fair they still will make mistakes. That's what the backup system is for!
My MCP server can control pretty much every aspect of the Hub, and has safeguards to disable/modify access for tools you don't want active, and has a read only mode as well. I'm constantly improving it and adding more features, if there's anything you want to see implemented that isn't already present freel free to open an issue on it. I've had a bit of a backlog lately but I usually address things fairly promptly, I've just been stuck on optimizing the RM tools the last few weeks.
King - thanks for your work on the MCP server. I installed it on both my hubs today and configured it to work with opencode locally. I then also configured it as a plug in for ChatGTP giving it read only access. Very nice job - a good step up from what I was doing.