Local AI integration

Can Comfy Desktop be integrated into the AI Girlfriends? It’s a Local image generation tool that runs a local server. It can be accessed by the browser via CORS (Cross-Origin Resource Sharing).

There are already sites that can use Local running Ai Tools via Comfy UI and LM Studio (My Setup) To run the AI completely local on the users pc generating Images and responses.

Like Ethumia (can use both) or Synsual (only LLMs)

fun.autoblow could be the front end only sending request keeping the costs low and enabling full AI functionality (and privacy cause it runs local).

Would be awesome if somebody could at least take a look at that. :grin:

We offer an open API. This is more a community-generated feature than something we would be able to develop and then fully support. I’ll link to it here in case any adventurous guys are interested in trying this. https://developers.autoblow.com/

You can do a lot of stuff locally once you get some basics sorted out. I was able to work with Claude Code and do this with Python; I had Claude do all the heavy lifting rather than use it as a guide. If you want to be 100% local, then you can to do the following (this is for Windows):

  1. Create a mobile hotspot.
  2. Connect your VacuGlide 2 to the hotspot.
  3. Edit your hosts file to redirect fun.autoblow.com traffic to your own web server (this is what I wrote in Python).
  4. Use your Python script to do whatever you want. Mine has a library of scripts similar to the OOTB scripts. I can also control it manually (speed up/down, engage valves).

I have not integrated it with ComfyUI for image generation though.

I know I just gave broad brush strokes above. If you have a Claude Code subscription, you can do everything you want by conversing with it in natural language.

I’m happy to share what I did, but I will see what Autoblow says first; maybe they don’t want people bypassing their site. Personally, I have no issue going to the site, but I try to keep as much local as possible. The only thing I have in my bag of tools that is not local is Elevel Labs for voices since I haven’t found a local voice generation model that get tone, emotion, etc. good.

We are happy that people do this. We put our full effort into developing these features that just work easily for everyone, like Video Sync. We improved it with VR Video Sync, the AI girlfriends and boyfriends, Autopilot, etc.We are totally dedicated to a “it just works for everyone” site that we give all owners for free.

When people want to do things like:

  • generate their own images for AI girlfriends
  • generate brand-new video clips that we have not generated
  • speak to it with their voice
  • Create a brand new AI girlfriend/boyfriend and give them machine access

These are features that we cannot support under a free model. That is exactly why we have opened our API for this type of use, and we are excited to see what people do with it.

The only limitation we have is on how many API calls you can do per minute. People who are developers know what it means to abuse the API calls, but perhaps we’ll get more novices who are using Claude Code and might use the API in a more abusive way without intending to do so. We address those cases as they occur, but people have been great so far.

1 Like

HI, Brian. When you say, “The only limitation we have is on how many API calls you can do per minute”, can you expand upon that? The code I run never leaves my network; no API calls are hitting your servers. My VaculGlide 2 connects to my hotspot and the hosts file is configured to redirect the API calls to my locally running Python app.

Hi, if this is not hitting our API servers then this would have no limitation. The firmware running on the device does not have any limitation in this regard but you mist be careful not to overload the firmware with too many requests as it might reset randomly.

1 Like

Cool privacy focused setup. How does auth between the device and your api server work?

It’s based on some keys and mac addresses. Our servers are a thin relay and routing system mostly to route commands sent to your token to the correct machine. Our devices persist a websocket connection to one of the closer regional servers.

But in case if you want to hijack the connection to your own setup this is not really required because you don’t need regional routing.