HooRii ClawStage – quirky AI that does Home Assistant

I wasn’t quite sure if I wanted to cover an AI-oriented product. I’m not a massive fan of the current AI craze and how everything must include LLM-based features, which we collectively decided to call “AI“. When HooRii reached out, I was about to say no, but then I gave it a chance as their display caught my attention. After a week of waiting, a brand new HooRii ClawStage was on my desk, ready to be plugged in.

Meet Miko – a cutsey virtual girl that thinks like GLaDOS*

*but you can personalise it to be whatever you want

HooRii ClawStage

In layman’s terms: an Alexa speaker with a fancy display, customisable personality and AI features that offer transparency via software (OpenClaw) and hardware (Raspberry Pi 5). The HooRii team wanted to create a soulful version of the assistants we all love (or hate) by changing how users interact with the hardware. With a strong focus on privacy and hardware transparency, HooRii offers a safe, fun way to use otherwise data-hungry LLMs.

The prototype gives me strong TARS from Interstellar vibes with its aluminium panels mixed with acrylic features and transparent display. At first, I thought that HooRii ClawStage took a stab at designing a holographic display; however, after closer investigation, it turns out that ClawStage features a high-definition transparent, touch-enabled display positioned in front of a light box to give that three-dimensional look. The cutsey anime characters may not be real holograms, but they look cool nonetheless.

The unit is equipped with a camera, stereo microphones with spatial awareness, and a speaker. The top part of the unit is mounted on a servo that has the ability to rotate the “head” and point it towards the sound (it’s impressively accurate at this).

The bottom part is less impressive, but offers practical features like: exposed Raspberry Pi 5 I/O and custom power ports, buttons and toggles. I appreciate the mute button disabling the microphone as well as the camera feed on ClawStage, as HooRii cares about the transparency, they shared the hardware implementation of the toggle on GitHub. There is a magnetic connector at the back that indicates various docs and expansions could be added post-launch, but as far as the I/O goes, HooRii ClawStage is pretty basic.

Rough waters ahead

Getting started with HooRii ClawStage was a little rough. I’m an Android user, and as such, I had no chance of getting through the initial setup without an iPhone. The HooRii team assured me the ecosystem would expand to include Android users after the Kickstarter campaign, but at the moment, there is much work to be done to streamline the setup process.

The team was happy to jump on the call and make sure my “loaned” iPhone was set the way it should be, and the HooRii ClawStage was ready to receive the updates and bind to my account. Once the setup was out of the way, I could start playing with HooRii ClawStage.

Please bear in mind that I had access to a pre-released hardware version of the device, and my impressions are based on that build. Post Kickstarter HooRii ClawStage may bring improvements and revised hardware.

iPhone and web interactions

As my phone was leased to me, I only had a quick look at Miko’s presence in the world’s most expensive mobile ecosystem. I like the animations showing the assistant entering and leaving the space, but other than setup, the app offered little extra in terms of features. It’s nice to use Miko on your phone, but opening the app each time to do it can get tiring really quickly.

What’s more fun is the AI personality builder available via https://workshop.hoorii.io/, where you can make your AI pretend to be anything you want (within the bounds of T&C) and http://localhost:18789 – the web dashboard of OpenClaw – with direct access to the same sarcastic abuse I have exposed myself to by defining toxic personality traits of my Miko!

The Dashboard is cool (this is my first time with OpenClaw) and shows a lot of promise, regardless of how you feel about HooRii ClawStage. I’m not going to pretend to be an expert on AI, but the dashboard is full of configuration tabs and options, turning the whole experience of customising your AI into a rabbit hole I’d like to explore more one day.

Shows promise

The most appealing part of the whole idea is adding a human language interaction to what otherwise would be a fancy Amazon Show with a much cuter face. HooRii ClawStage is great at recognising my sentences, and all interactions sound very natural (if it wasn’t for the fact that it took on the personality of GladOS – don’t worry, you can completely change the personality of your assistant).

When connected to an LLM model, HooRii ClawStage can look up the information online and within the context of the conversation to craft clever (and in my case, sarcastic) replies. The language sounds impressively natural through the built-in speaker. I like how the context of each interaction always stays in the thread, and the assistant is very keen to remind you of the past interactions – no matter how infuriating Miko gets each time.

For the first time, I used natural language to interact with my smart home. I linked up my Home Assistant instance with ClawStage and minutes after I could command my home without being restricted to pre-defined commands.

Brilliant!

Hey, Miko – what’s the solar energy production for today? Is it worth getting my laundry done? – was’t meet with Alexa having a fit trying to figure out what to do with that sentence. Instead, Miko meticulously checked the solar power predictions, confirmed the weather forcast and confirmed that I should get about 2000W of power available to me through most of my day, and it’s sunny enough to get my laundry.

Encouraged by this, I took Miko to my bedroom for the movie night. I wanted to see how ClawStage would fare when used with camera inference. I asked questions about what’s on the screen, asked her to recognise actors and suggested TV recommendations based on my preferences. While Miko was knowledgeable, she wasn’t prone to mistakes. She confused actors when trying to guess the name based on a single (in-character) shot from my TV. This is likely due to the limited resolution of the built-in camera and the distance to my TV (approx 2.5m). Other than the latency in responses (caused by querying the cloud-based LLM and taking pictures), the interaction was actually fun.

The most impressive feature of the HooRii ClawStage isn’t really advertised loud enough. An AI-powered device is already running a Home Assistant instance in the background, but as I already own mine, I didn’t want to recreate it from scratch for the purpose of testing. To set everything up and query my smart home, all I had to do was:

Within seconds, I was able to query my server with natural language commands to get status updates about my solar panels or home automations. This is where LLM-based AI shines – understanding human speech rather than relying on strict and often hard-to-remember commands. The ease of configuration and retrieving information in human format was simply shocking. Using AI to configure services on your server makes perfect sense. Nobody likes configuring files and following countless tutorials.

The double-edged sword – the AI

The biggest negative of HooRii ClawStage is the time taken to compose the replies. At the current speeds, each query takes several seconds to be prepared before it is spoken back to me. The fan’s loud whine and background sounds also contribute towards Miko (the default assistant personality) stalling and waiting for the sounds to fade before submitting the request to your fav LLM. I’d appreciate it if more work were done on the noise suppression in the microphone.

This isn’t as bothersome when I use AI web clients, as I’m used to pages taking their time to load the answers, but when imitating human speech interactions, unnaturally long pauses are very immersion-breaking. I had to turn on captions to see if things are working.

Costly fun

After a month of poking about with voice commands and diving into OpenClaw dashboard I discovered a horryfying reality of using cloud LLMs. Simple voice interactions won’t use up many tokens, but as Miko remembers the past, even the cost of voice queries become significant. This gets worse when sophisticated queries are used like: image inference, API calls and self troubleshooting.

The open-source nature of the framework gives the illusion of free AI interactions, everything HooRii ClawStage does uses precious AI credits. Depending on your subscription level to services like ChatGPT, DeepSeek or Claude, the cost may vary. Having advanced interactions can burn through the credit allowance relatively quickly, while generating new assistants from scratch will cost you easily a couple of hundred AI credits. OpenClaw allows you to link up with a local LLM hosted on a faster-than-Raspberry Pi computer, but then expect to learn more about LLMs and how to set one up running locally on a computer with a beefy GPU to speed things up.

In my time with Hoorii ClawStage, only using it sporadically for testing and shooting showcases, I burned through 27M tokens. That’s 27 000 000 tokens used across all my activities across several sessions. AI is a capable tool, but comes at rather high cost at times especially when dealing with complicated toolchain calls and image inference. While usage limits can be set in OpenClaw, it’s worth monitoring the usage to be up to date with the cost.

Teething problems

AI and LLMs are an emerging technologies. Everything gets developped and updated at rapid pace at the moment, while developers are trying to make better models and easier interfaces for customers to use. Hoorii ClawStage relays on an entire toolchain of software and API to work and couple of times, I experienced unexplained issues caused by updates.

I was amused and impressed by ClawStage’s ability to troubleshoot itself and evern fix some of the issues (Miko tried clever CLI commands to see what was wrong – something I’d struggle to replicate with the same scope and timeframe), but it’s too early to call the software “mature” when toolchain gets broken due to incompatible/unsupported software versions.

HooRii ClawStage just got funded on Kickstarter. That will help address some of the issues listed in this review. It’s a passion project, and after speaking to the team, I do not doubt that they care about what they have created. In a world where everyone uses AI to embed it in yet another customer service chat, the HooRii team bets on fun and silly interactions. One thing I know for sure, there is work to be done ahead to turn HooRii ClawStage into a more reliable device that doesn’t rely as much on the user to fix things up.

The question is: Will this bet pay off?

Final thoughts

HooRii ClawStage is an experiment. It will have just as many admirers as people who see no reason for the device to exist. If you treat it as a final product and expect it to replace the brain of your home and become the go-to interaction, I feel that you may get disappointed with the lack of polish. If you are, however, keen to try to squeeze everything you can from this little AI box and enable some rather impressive (given the shortcomings) skills, then HooRii ClawStage bring the power of OpenClaw, transparency and hardware that enables more than simple CLI interactions. Let me know if you like or hate the AI experiments in this Reddit thread.

🆓📈💵 – See the transparency note for details.