What the title says. It would be nice to have to option to use own local AI instead (like LM studio). This would help with the servers sometimes being very slow and sometimes not responding
2 Likes
To be honest here, this is a pretty niche use case for us to develop and maintain. I’ll bet you that if we offered this, we would have fewer than 20 people out of thousands using it. Unfortunately, it’s not something we can devote our resources to.We do have an API which allows people to set up things like this who requires them.
There are sometimes limitations on the AI models themselves and how long they take to respond, but the technology on this is evolving, and we are updating this weekly.
1 Like
I am getting a ‘Provider Error’ relatively often and therefore no answer. Six inputs from me and no answer. Otherwise, the AI girlfriend function is something I really enjoy.
Thanks. We will let our engineers know about this and address it tomorrow. If they need any specific questions answered, I will reach out to you by private message here.
This is very, very new technology, and so such things can happen.But I do appreciate you letting us know about it because that is also how we find out that we need to fix it.
Update: apparently one of the AI models we use suddenly started providing 10% of the responses with that error because it was refusing to do adult content.Even though for the last month or two it had been excellent at providing adult content without issue, it suddenly decided it didn’t like adult content.We have already fixed this as of last night, I just learned.
I was able to get my local AI to control my VG2. First step was feeding claude the link to the VG2 API doc site and having it create an mcp server. Then I containerized that and deployed it (feeding the device token in via environment variable, but you could hardcode it). I’m using SillyTavern as my frontend and had to use bmen25124’s SillyTavern-MCP-Server and client. Install and enable that, then add the server.
For me, it works okay. Pretty good latency, the only problem I’ve had is that I suck at writing character cards and system prompts to get them to use it as effectively as the AI Girlfriends feature.
1 Like
wow a video of this working would inspire others if you can swing it
1 Like
(Video link since the forum wants to embed it - Watch PXL_20260806_003815958 | Streamable )
Here’s the code if anyone wants it - Gofile - Cloud Storage Made Simple
Use Docker to create an image from the dockerfile, then run it using docker. I’m not sure on the wiring since I use kubernetes at home rather than docker, but you should be able to feed that to an llm and get a walkthrough on how to set it up on your system.
This was flagged as spam for some reason by discourse, so I unflagged it for you. This is awesome. I really appreciate you posting this. I’m sure this will inspire more than a few people.
1 Like
Thanks. I was wondering what it was that triggered the hold. Figured it was probably the file sharing domains might be on an automatic hold list.
Would you all be willing to provide the system prompts and/or character cards for the AI girlfriends? As I mentioned up-thread, I haven’t written a prompt for a character to control it as well as yours.
Also, I don’t really want to get into your finances, but you guys provide a lot of content for free ( you know, after the initial purchase ). I’m thinking the LLM costs alone wont be sustainable once the providers start increasing token prices.
Have you all considered providing a package like you do for the 3d print files that contains the components for running the same characters but locally? I could probably package everything together into a docker image, and you all are more than welcome to the mcp code I had claude write. That might help those of us who are a bit more privacy focused and potentially lower your token costs.
No, I don’t think the guys will share our prompts. Im sure there are a lot of tutorials online for this if you search, or you can even ask the LLM to write one for you to achieve a certain result.
The LLM costs are ok, and we don’t anticipate they will be an issue in the future.
We do not consider offering information/support/components for running them locally. This is a part of our service we provide to machine owners.
Fair enough, thanks for responding
We spend a lot of time testing and improving our AI agent harness and we optimize it to lower the costs all the time. There is a lot of work to make sure the AI Agent harness works the way it currently does so most likely sharing our AI cards would not really get you the same experience as it would on our own platform.
We are going to use an improved model that should offer an even better experience in the next few weeks.
Naturally.
I’ve only spent about an hour messing around with adding in system prompts for the vg-mcp that I linked earlier. I’m primarily interested in figuring out how to create something that can be applied to a frontend like sillytavern that works regardless of character card and adjusts the machine based on the content of the chat. I only though about promoting, not about trying to wrap that functionality in a harness. Thanks for the idea!