HA-Architect: An AI Assistant that actually knows your entities (plus Automation Library)

[Phynix]

Hi everyone,

I wanted to share a project I’ve been working on in my spare time and see if there’s any interest in the community for something like this.

The Problem I love using tools like ChatGPT to help with complex YAML automations. However, one thing has always annoyed me:

The AI doesn’t know my smart home.

I either have to manually copy-paste my entity IDs into the prompt, or the AI starts hallucinating devices that don’t exist in my setup. I couldn’t find a tool that solves this workflow smoothly while keeping the context of my actual devices.

The Solution: HA-Architect So, I started building my own Add-on: HA-Architect.

The concept is simple: It is a chat interface living directly inside Home Assistant, but it has read/write access to your configuration and context.

Key Features (see screenshots):

Context Aware & Filters The tool scans your entities and areas. You can simply tell the chat (or use the UI filters) that you want to automate the “Living Room.” The tool then feeds the AI only the relevant devices from that area. This saves tokens and ensures the AI uses the correct Entity IDs.

Automation Library This is a major feature for me. You can save the generated automations directly in the Add-on. This creates a clean library where you can manage, edit, download, or delete your AI-generated scripts. It keeps your chat history clean and your automations organized.

Direct Installation If the AI suggests an automation and it looks good, there is an “Install in HA” button. No more copy-pasting code into automations.yaml.

Complex Logic It handles complex requests easily (e.g., “When Plex starts on Apple TV, dim the lights, but only if it’s after sunset”).

My Question to you: Is this something you would use? Do you also miss a “Copilot” that actually knows your context and helps manage the files?

I am currently still deep in development, but if there is enough interest, I would love to get some beta testers on board later on.

Let me know what you think!

Cheers, [Your Name]

[IOT7712]

@Phynix

living directly inside Home Assistant, but it has read/write access to your configuration and context.

@Phynix

Cheers, [Your Name]

Looks like you’ve been assimilated already.

[Phynix]

Yes, I was a bit lazy and had my thoughts on the subject summarized

[IOT7712]

@Phynix

Yes, I was a bit lazy and had my thoughts on the subject summarized

From a conceptual aspect, abandoning the vendor ecosystem safety to go independent and cloud free with HomeAssistant, and then hand over the keys to your universe to a third party ecosystem that is trained on out-of-date information sounds like it may not be a valid choice.

Should Apple (using them as an example as you seem to be tapping into their ecosystem) designers and programmers offer up their entire source code and intellectual property to your artificial intelligence engine, on an ongoing basis, then I would possibly give it a go, otherwise, a very, VERY hard NO.

The home automation industry is moving at enormous speed. Unless artificial intelligence is fully involved in the design and implementation, it will always be behind and playing catchup. Do you want obsolete and possibly incorrect choices guiding your lifestyle behind the scenes, deeply involved in controlling your life?

Do you think your artificial intelligence engine knows more than the vendor? If yes, go that route, otherwise stick with the vendor cloud based system, and the comfort and security implications that involves.

Selling your soul to another devil is still fraught with danger. Being mentally lazy will probably not end well.

[Phynix]

That is fair skepticism! I am a big proponent of “Local Control” myself, which is why I built this as a local Add-on rather than a cloud service.

However, I think there might be a misunderstanding of the actual architecture regarding your concerns about outdated data and security.

The “outdated information” issue is exactly what I designed this tool to solve. It doesn’t rely on the AI’s training data to know your devices. Instead, the backend scans your live registry, stores the current state in a local SQLite database, and injects that real-time context directly into the prompt. The AI isn’t guessing based on old Apple specs; it is processing the strict schema and entity IDs that my system just handed to it effectively in real-time.

Regarding the “keys to the universe”: The AI has zero direct control over the house. It acts purely as a code generator running in a secure container via standard HA Ingress. Nothing happens until I manually review the generated YAML in the “Automations Manager” and hit the install button myself. It drafts the blueprint, but I remain the builder who has to approve it.

As for being “lazy”…I see it more as removing friction. I know how to write YAML, but I don’t enjoy manually looking up whether a specific motion sensor is binary_sensor.lumi_motion_v2 or _v3. The tool handles the syntax and ID lookups so I can focus on the logic. It’s definitely not for everyone, but for those who want to combine LLMs with their actual local context, this bridges that gap safely.

[IOT7712]

Keep us posted on progress, warts and all. Wins and losses. [Unfiltered]

[Phynix]

Of course, I will! I really appreciate the challenge. I think honest skepticism is so important for keeping the ecosystem healthy. I’m going to be really open about the things that might go wrong, as well as the things that go right. Thank you so much for keeping an open mind!

[IOT7712]

Remember, not cutting edge, bleeding edge!
I’ve made enough mistakes in life to know to sit back and learn from others making them first.
Yes, stand on the shoulders of giants, looking to the horizon, but be aware of how far you have to jump down to run away as well.

[Scryan7371]

I would love to give it a try when it is ready. I think it could be super useful for very complex tasks and to refine existing tasks that are close but not perfect

[IOT7712]

@Scryan7371

refine existing tasks that are close but not perfect

Observing posts on this forum would tend to indicate the very opposite, AI LLMs offering a nearly finished solution, usually wrong, however only needing a few human touches to correct.

Those human touches often require extensive reading of vendor documentation along with understanding, which if done first, could have prevented enormous confusion and wasted time.

[AshaiRey]

@Phynix

but it has read/write access to your configuration and context.

Well I stopped at the very first moment that I read this…

No sociopatic AI with amnesia or any other AI for that matter ever will be able to write to my config. No need too also. I am already able to make catastrofic mistakes myself. No need for an AI to do that for me.

[Dorschdne (Thorsten)]

I for one would be happy to try this! I don’t have very complicated automations yet since I couldn’t find the time to really get into the subject of yaml programming. I have only recently switched from using FHEM and OpenHab and would appreciate the help from an AI assistant!