Page 1
Page 1
Started By
Message

Anyone using Home Assistant AI agent?

Posted on 3/2/26 at 6:22 pm
Posted by Dallaswho
Texas
Member since Dec 2023
3562 posts
Posted on 3/2/26 at 6:22 pm
This thing is so damn awesome but there is a HUGE problem. The speech to text listener, at least on the iOS app is completely useless unless you’re in a quiet room by yourself. It’s not that most attached STT models can’t easily cancel background noise but the app itself just keeps listening and listening until things are dead silent. That’s never the case around me. Completely useless. Still the agent is so good it’s worth typing.
This post was edited on 3/2/26 at 6:22 pm
Posted by LemmyLives
Texas
Member since Mar 2019
15386 posts
Posted on 3/2/26 at 6:24 pm to
What did you use to get started with it? I saw your other HA post a few hours ago and am curious.
Posted by Dallaswho
Texas
Member since Dec 2023
3562 posts
Posted on 3/2/26 at 6:37 pm to
Qwen3-30b locally but you could probably get with something smaller. Openrouter has a bunch of free or super low cost options.

Used extended OpenAI integration for LLM api and Wyoming for TTS and STT.

Setup is pretty easy but function has to be pretty much perfect and prompt should explain what you want it to do.just use chat.qwen.ai for that. It fixed everything for me when Gemini couldn’t.

Apparently if you use an esp32 device for STT there are options for speech cut off sensitivity, but not if you’re using the iOS app. They just need to go to a button or something if they can’t get it right.
Posted by Lazy But Talented
Member since Aug 2011
15063 posts
Posted on 3/2/26 at 10:47 pm to
Will have to look into. Just starting in my HA journey. Used Claude code to turn my bricked Surface Pro into an Ubuntu server.
This post was edited on 3/2/26 at 10:48 pm
Posted by Dallaswho
Texas
Member since Dec 2023
3562 posts
Posted on 3/8/26 at 2:42 am to
Been drinking beer tonight but also trying to move home assistant bot to openrouter for better speed. I must have lucked out my first guess because all of these way more powerful models are absolutely shite at following directions and calling tools. The OAI GPT models are literally 0%. Bigger Apache2/MIT models are maybe 60% while the old trusty qwen3-30b-instruct(even at q4) gets the call right basically 100% of the time. Still going to try Qwen3.5-35b but this is ridiculous.
Posted by Dallaswho
Texas
Member since Dec 2023
3562 posts
Posted on 3/13/26 at 11:19 pm to
My bot quit working again tonight. Context drift. God these tools suck arse. Need automation to clear conversation history every few hours because it won’t do it on its own. This is “extended OpenAI conversation” by the way.

Also qwen3.5 flash for about a penny or less per day is about perfect. 2-3 second tool call and works every time.
Posted by Zach Lee To Amp Hill
New Orleans
Member since Mar 2016
4928 posts
Posted on 3/14/26 at 8:43 am to
I’m using openclaw and it’s great for this.
Posted by Dallaswho
Texas
Member since Dec 2023
3562 posts
Posted on 3/14/26 at 4:17 pm to
How on earth would openclaw be good for this? It’s going to cost 50x more and have ~7x the latency just in overhead. I’m only sending about 50 entities and a 6 part function.
Posted by CAD703X
Liberty Island
Member since Jul 2008
93032 posts
Posted on 3/14/26 at 5:56 pm to
The more I get into HA the more this thing feels like a cult.

The weird dude on eBay printing white boxes for every possible type of sensor on the known universe confirmed it.

But it's insane everything from cheap rando Chinese devices to my Bluetooth coffee machine to xsense smoke detector system to the damn dishwasher and even the cheapo ups hooked up to my PC and aside from a few integration issues and having to redo a few Chinese devices in tuya so I can delete those 15 stupid apps the only real issue is my eufy locks. I wish I had known before I installed them everywhere in my house they didn't play well with ANY integration outside of the native app.

Oh and I have big box of rgbw lightbulbs to redo all my outdoor path lighting and it had the word MATTER on the side so before I installed them as wifi asked Claude and he started singing the praises of matter and the fact I can set them for local control and skip the wifi crap altogether.

This is a serious time sucker
This post was edited on 3/14/26 at 5:58 pm
Posted by CAD703X
Liberty Island
Member since Jul 2008
93032 posts
Posted on 3/14/26 at 7:00 pm to
Good Lord I just realized my Fitbit also has a plug in

I may add a health dashboard too because why the frick not.

"Hey HA if my steps go over 10k play We Are The Champions and make all the lights in the house flash"
Posted by Zach Lee To Amp Hill
New Orleans
Member since Mar 2016
4928 posts
Posted on 3/15/26 at 8:17 am to
I got a free one week Claude pro subscription, used it to write all the automations and coding and shite and then switched to Gemini 2.5 flash for the tool calls and shite after that's done. It's extremely cheap.
Posted by Dallaswho
Texas
Member since Dec 2023
3562 posts
Posted on 3/15/26 at 11:54 am to
So instead of pre-filling entities, examples and one function, you give it an entire openclaw ecosystem? That’s either going to add a ton of latency and/or add errors via look-ups. There no way around those things. Plus the wife is going to complain if it messes up one single command even though she’s tolerant of Alexa screwing up 10 in a row.
Posted by Dallaswho
Texas
Member since Dec 2023
3562 posts
Posted on 3/15/26 at 11:59 am to
Ok now I’m getting the itch. I mainly use for lights, alarms, and access.

Wait until you get into the frigate rabbit hole. Everything you can see becomes a sensor.
first pageprev pagePage 1 of 1Next pagelast page
refresh

Back to top
logoFollow TigerDroppings for LSU Football News
Follow us on X, Facebook and Instagram to get the latest updates on LSU Football and Recruiting.

FacebookXInstagram