I wanna put my two cents on this as someone who fucks with them semi-regularly:
Local LLM hosting faces two problems: limited context size and power requirements.
Basically, in order to have a good model you need power; the current mid tier open source LLMs require the tippy top of what is available to the customers, with the best models (the ones with 70B+ training prompts) basically being locked behind whole operations, like those subscription services you see around or huge industrial tier computer labs like OpenAI or Google.
At the same time, the AI can only 'remember' a limited amount of stuff before it starts making shit up. So if you're playing, lets say an AI based LT, and you enslaved a guy named Pete but then went to do other shit, by the time you returned to Pete the AI does not remember it anymore. There's currently a few models that address this; OAI's GPT models and Google Gemini, and they do so by having a dynamic context limit, which basically reads the entire thing and generates keywords for certain things that happen in context. This requires an immense amount of processing power and no singular person alive that isn't some arabic prince can afford it.
So then you're faced with the alternative of always online porn games, which need either a subscription model or to have sponsors keeping them afloat, and as we know, people with money think porn is icky - especially when it comes to uncensored AI. Plus the privacy concerns of sending your smut to someone you don't know who may be selling your shit to third parties to keep their project alive. At any rate, in both cases, no computer savvy writer doing their first project in Godot or Java is going to be able to pull such a thing out.
So yeah, unless we find a way to have futuretech GPUs with unlimited VRAM and CPUs with unlimited processing power, or some way to superquantize AIs so they can run anywhere, we're not getting that anytime soon.