SillyTavern is genuinely excellent software. Open source, 25,000+ GitHub stars, actively maintained, connects to any model backend you want, and gives you total control over everything. The TavernCard character format — character data embedded as JSON inside a PNG — is the closest thing the AI fiction community has to a universal standard. We adopted it ourselves.

We're not going to pretend SillyTavern is bad so we look better. That would be dishonest and you'd see through it anyway. What we will do is explain why we built Underfiction when SillyTavern already exists, and who each tool is for.

The setup question

SillyTavern is a self-hosted Node.js application. You clone a repo, run a server, and connect a backend — a local model via KoboldAI, a cloud API, or something like Oobabooga for running open-source models on your GPU. If you know what all those words mean, you'll be fine. Setup takes fifteen minutes to an hour.

If you don't know what those words mean, you're looking at a weekend. And that's the gap we built Underfiction to fill. Open a browser, sign in, write. No servers, no API keys, no backends.

What each does better

Models and prose quality

SillyTavern is model-agnostic. Connect it to Claude, GPT, Gemini, a local Llama, whatever you want. The output quality depends entirely on which model you pick and how you prompt it. With the right setup, SillyTavern can produce incredible output. With a default system prompt and a mid-tier local model, it'll be mediocre.

Underfiction has a curated model list — Opus, GPT, Gemini, and Grok — and system prompts tuned specifically for literary fiction. You pick a model and a narration style. The floor is high. The ceiling is lower than a perfectly configured SillyTavern setup, because you don't control the prompts. (You can edit the system prompt in advanced settings, but it's not the same level of granularity.) One thing worth noting: because these models run through Venice rather than through the official APIs, they're noticeably more willing to follow creative direction. The same model that hedges or refuses via the official API will write the scene as directed through Venice.

Long stories

This is where the products diverge most. SillyTavern handles long conversations with lorebooks, world info entries, and various truncation strategies. It's powerful and flexible, but you're the one managing context. When the window fills up, you decide what gets cut.

Underfiction structures stories as scenes. Each scene has its own context. When a scene ends, it gets summarized automatically and the summary carries forward. Long scenes compress in the background. You never think about context management. The tradeoff is you get less control over exactly what's in the context window at any given moment.

Privacy

SillyTavern with a local model is the gold standard for privacy. Nothing leaves your machine. Period. If you're connecting to a cloud API, though, your prompts go to that provider and their privacy policies apply.

Underfiction stores stories locally and routes inference through Venice, which strips your identity from prompts. It's not as private as a fully local setup, but it's structurally more private than most hosted alternatives. If you want to sync stories across devices, Underfiction offers optional Story Sync with server-side encryption at rest.

So which one?

If you enjoy building your own setup, want total control, or need to run local models — SillyTavern. It's not even close. It's the best tool in its category.

If you want to open a tab and start writing literary prose with frontier models — Underfiction. We built it because we wanted SillyTavern's output quality without SillyTavern's setup process.

They're not mutually exclusive. Some people use both. We do.