No, I think it’s an actual bug. It should be patched, just update.
Hehe, best support. :) Glad to hear and i don’t need to potentialy mess up my system. 
Edit: Happy to say, it is indeed fixed.
Edit2: On thing heals, another breaks … currently getting
“The local model returned an unreadable response. Check Ollama model status and try again.”
Even tho for half a second i can see that a valid massage was send that gets changed to that errortext afterwards.
That’s the local model giving a bad response (not formatted to prompt requirements). Look in the model memory section and see if it has anything weird in there and delete it. Or try another model. I still haven’t tested a bunch of uncensored models, there are probably better options than what is default right now.
There are more options on huggingface, but those you would need to load into Ollama manually then add the name to settings. If it’s on Ollama, StrokeGPT can just download it for you.
Let me know what model you were using and send anything that may have confused the model in memory.
I just updated the app to send the message anyway and warn about the malformed response. I haven’t been able to replicate a malformed response even when literally asking the LLM to, the current prompt seems reliable.
Reinstalling Ollama was the fix.
I just pushed some updates to motion handling. It was not respecting higher speeds, that should be fixed.
I wonder, is it much work to put up a “release” and a “test” at the same time? Since yesterdays build is very stable and works well ; So it would be nice to test without potentially bricking the App. ;)
Yes, I do plan to switch to having a stable and a nightly branch at some point. Once I am done making major changes all the time.
Are you saying there are new issues with the current version?
Nono, i haven’t found time to test yet, maybe later i can give feedback.
Found some time to look into the current version. Works mostly well, but after a few minutes the normaly clear and fluid motions become jittery again (like a few versions ago) ; I noticed it is most pronounced when it is “morphing”.
Update and try again.
Was able to test again, so far it works great, yet the issue still persists. I do believe it might be part issue that the ai does spam commands so that the handyfeeling interface starts blocking/(timing out) it, if this is a thing, which i am sure it is (because having no protection would be silly).
It also gets wonkey if the ai changes modes/patterns on my end (chat mode->edging mode or trying to use a specific pattern) to the level where it fully breaks because the commands appear to glitch out. Wonder if this is only on my end.
But since i am not that informed, do you think you could make a direct bluetooth connection to the handy work , since the Handy HSP Bluetooth seems extremly stable imo.
I pushed an update that may help with mode transitions but I think it may be a different issue. Is this just when using chat or is this when using freestyle or the edge/milk buttons?
I have been thinking about implementing bluetooth, though idk if it will really improve this. You might see an early implementation out today.
In Chat-Mode usually triggered by an input causing the AI to switch into Edge/milk mode is where it completly fails without recover (besides a Stop-Command).
And on the bluetooth side, i am not sure aswell, but since bluetooth at least in my experience is generally way more stable esp with many commands send / fast switches it tends to be stronger for applications that don’t “just play a script”.
Really appreciate the work done on this! Been enjoying it!
Just wanted to chime in and request that someone smarter than me put this in a docker container at some point. I’d really like to host this on my server and use my main GPU…for other simultaneous…activities
Hey, i want to contribute into this project, found it awesome, i have a RTX 5070 and The Handy how can i help in this?
I gave this a very brief spin. It’s an interesting project!
A few comments:
Thumbs up for buttplug.io, the Handy is not my favorite.
Consider adding an animated avatar? I use OpenLLMVtuber (Live2D model) for general chat, and the avatar adds a lot to the experience.
Speaking of that app, it does some clever, I assume, things to TTS; running Sherpa-onnx in CPU mode makes for quite a smooth experience - though it will slow down a little when given a wall of text.
You asked for some Ollama model recommendations, here are two:
draganis/vanessa
14GB iirc, quite fun if a little limited.
Fermi/Cydonia-24B-v4.3-heretic-vision:Q4_K_M
with 128k context you don’t get much change out of 24GB VRAM, but you do get 1000+ word stories that remain coherent and stay on topic, all be it with a little prompting. Paste 20K of text into the prompt, and you can have a text reader with attitude - it tells me it could drive your interface, ymmv. Also, if you ask it to dial back on the smut, you can have some interesting and amusing conversations.
N.B. As always, the prompt(s) are key.
Reporting bugs, testing different models on Ollama to see which work better for you and reporting here, writing custom prompts and sharing them. If you are willing to use a coding agent, setting up Codex or Claude Code to work on the app itself also works! You don’t need to know how to code, just need to have decent critical thinking/writing skills and basic computer literacy to know what to ask for and how to ask for it.
Any superfluous features like that would be for later, the app is still pretty rough imo. Adding more technical debt is not a good idea, I’ll just get burnt out. It’d be great if someone else could contribute to the repo though!
I am mostly focused on making it seem like the LLM does exactly what you are requesting. At the moment it isn’t there. I have some ideas to enhance immersion but that mainly revolves around pulling you away from the chat interface, not making you stare at it the whole time.
Thanks for the model recommendations, I will try the vanessa one (actually only 4.5GB), I only have a 5070 Ti so I cannot test the larger models myself.