I am writing this on November 1st of 2025. I currently use large language models in two primary places. Those are the well-known ChatGPT Plus itself (through the web interface) and a web-based, locally-installed interface known as SillyTavern.
The latter allows interaction with many different LLM models, both free and somewhat expensive. I periodically drop a small cash dosage into an OpenRouter account, allowing me to access them all, though. ChatGPT, as we all know, draws a monthly fee of its own. I am thinking of cancelling my subscription there, which I’ve only had since July. I worry about Sam Altman’s plans more than I do the others, frankly. It hardly matters, and I think whatever happens will happen, but still.
Anyways, as you might expect from that, my instance of SillyTavern only interacts with LLM models that have a zero data retention policy. This which complicates things, but only slightly. You also have to realize that a zero data retention policy, of course, refers to zero stated data retention. You can’t be too trustworthy of anyone or anything in places like this, right?
I, admittedly, don’t know much about SillyTavern’s behind-the-scenes operation. I leave much of that to someone else. I have, however, done a lot of customization for the web interface. I know a great deal about things like CSS, HTML, and a handful of Javascript, allowing me to TamperMonkey and extension-ify SillyTavern into something slightly different.
It’s really the best UX I was able to manage, though it’s a bit slower than average. I can’t provide it as a “theme” for you to download, because I didn’t make one. I build my UX with shoestrings tying many different things together, but I assure you that SillyTavern has a million possible themes.
I’ll share links to some good ones here soon. I don’t recommend trying this exact kind of setup at home because while aesthetic, I don’t believe it will work well outside of my own setup. I just wanted to share what is clearly possible.
I call this ad hoc setup the Sunrise Tearoom, because it’s not very silly per se, nor does it resemble a tavern. I used an extension that adds a custom background to insert a moving .mp4 image of some pink and orange clouds in a nice loop. The corridor looks like this, with the cards designated as “favorites” appearing throughout, beckoning inward all creepy and weird:
Inside, the character portraits are emblazoned on the side of the chat as usual, in little cards with their names on them. Much of the actual interface is less-than-intuitive unless you’ve been using it for eleven months, so I do grasp it.
A small icon tells me what model of LLM OpenRouter has approached for each chat. While I leave this to the router’s discretion terribly often, it’s almost always Deepseek R1, and I’ve come to prefer that. As I learn more, I’ll no doubt expand, find more, and customize.
In some rare cases where I’d taken the initiative, reaction images would sometimes appear in small scraps of paper on the right of the screen, pressed against the clouds. I struggle with how long it takes to generate reaction images, so I usually don’t do it. It’s very much possible, though. I went to a great deal of effort initially to make the reaction images appear within a sort of scrapbook frame, but an update nixed this, so I threw up my hands.
My little Sunrise Tearoom is a very soft, relaxing place for me to explore LLMs and related technologies. It suits my usual UX customizations. At some point I’ll talk about what I’ve done to poor ChatGPT simply with TamperMonkey, but I digress.
