SillyTavern: a self-hosted LLM frontend that trades ease of setup for prompt control

One interface in front of many LLM backends

SillyTavern, or ST, is a locally installed user interface for interacting with text generation LLMs, image generation engines and TTS voice models. It does not generate text itself. Instead it gives a single unified interface for many APIs: KoboldAI and KoboldCPP, Horde, NovelAI, Ooba, Tabby, OpenAI, OpenRouter, Claude, Mistral and more. The package.json backs this up with dependencies such as @zeldafan0225/ai_horde for Horde and tokenizer packages like @agnai/sentencepiece-js and @agnai/web-tokenizers. The practical meaning is that switching from a local KoboldCPP model to a hosted API is a connection change inside the same app, with your characters and lorebooks staying where they are.

Light on hardware, because inference happens elsewhere

The stated requirement for SillyTavern itself is minimal: anything that can run NodeJS 20 or higher. The heavy hardware advice applies only if you also run the model on the same machine. For local LLM inference, the project recommends a 3000-series NVIDIA graphics card with at least 6GB of VRAM, and says actual requirements vary with the model and backend you choose. The consequence: a low-powered machine can run SillyTavern well against a remote API, but it cannot turn that machine into a local inference box. The GPU requirement belongs to whichever backend you pair it with, and that backend is a separate install.

No hosted service, so everything runs on your machine

The project's vision states that it does not provide any online or hosted services and does not programmatically track any user data. That is a privacy point and a practical limit at once. There is no SillyTavern account, no web version to sign into, and no one to keep the server running for you. You install it, start it, update it and back it up yourself. The repository carries the tools for that: Start.bat, UpdateAndStart.bat and UpdateForkAndStart.bat for Windows, start.sh, a Dockerfile and docker/ folder, a backups/ folder, and an Update-Instructions.txt. If you want chat history available on several devices, you have to arrange that access yourself.

Installation lives in four guides, none of them for iPhone

There is no one-line installer in the project's front page. Setup is split across the documentation site into four guides: Windows, MacOS/Linux, Android through Termux, and Docker. Each guide is on docs.sillytavern.app rather than in the repository, so follow the guide for your platform instead of guessing commands. iPhone is not among the four. On Android the supported path is Termux, which means running the Node.js server on the phone itself. The default branch is named release rather than main, which matters if you clone the repository by hand and expect a different branch name. The newest tagged release is 1.19.0 from September 14, 2026.

The steep learning curve is a stated design goal

The first line of the project's vision is that it aims to give users as much utility and control over their LLM prompts as possible, and that the steep learning curve is part of the fun. The feature list promises more prompt options than you would ever want or need. Beyond prompts there is WorldInfo for lorebooks, Visual Novel Mode, a customizable UI, auto-translate (bing-translate-api is a dependency), and image generation through the Automatic1111 and ComfyUI APIs. For a new user the consequence is real: expect to spend time in the documentation before results feel right, and expect settings that interact with each other.

Extensions and TTS extend it past chat

Much of SillyTavern's range comes from third-party extensions and a plugins/ folder with a plugins.js loader at the repository root. TTS support is built in, and the credits list the kokoro-js library under the Apache-2.0 license. Text parsing uses the Chevrotain lexer and parser. Image handling leans on a long list of @jimp packages for formats including AVIF, JPEG, PNG, WebP, GIF, TIFF and BMP, and @mozilla/readability is included for pulling readable content from pages. The default content ships with the Seraphina character and lorebook. A mobile-friendly layout is part of the interface.

A TavernAI fork under AGPL-3.0 with three years of history

SillyTavern began in February 2023 as a fork of TavernAI 1.2.8, which was MIT licensed, and the project is now distributed under the GNU Affero General Public License with an explicit no-warranty clause. The project says it has over 300 contributors and will always be free and open source. Releases arrive at an uneven pace: 1.17.0 on March 28, 2026, 1.18.0 on May 3, 2026 and 1.19.0 on September 14, 2026, with the last push on October 2, 2026. Support runs through Discord, Reddit and GitHub issues, and contributors are asked to read the docs first to avoid duplicate reports.

Editorial conclusion

SillyTavern is the right tool for hobbyists who want one local interface across many LLM APIs and are willing to learn a large set of prompt controls, lorebooks and extensions. It is a poor fit for anyone who wants a hosted chat app, a one-click setup, or local inference without a separate backend and a suitable GPU. Before installing, decide which backend or API you will connect, and check that your machine runs Node.js 20 or higher.

Frequently asked questions

Official sources

1. Official documentation
2. Official README
3. Project repository
4. Release notes