SillyTavern

SillyTavern is a self-hosted, browser-based frontend for interacting with text-generation LLMs, image-generation engines, and TTS models. Originally a fork of TavernAI, it has evolved into the de-facto power-user interface for character-driven roleplay, collaborative fiction, and long-form conversational writing. It runs locally (Node.js) or in Docker and connects out to whichever backend you point it at: OpenAI, Anthropic Claude, Google Gemini, Mistral, OpenRouter, KoboldAI/KoboldCpp, Text-Generation-WebUI (Ooba), Tabby, llama.cpp, NovelAI, AI Horde, and any OpenAI-compatible endpoint. Because it is only a UI, no data flows through a SillyTavern-owned server — your prompts and chat history stay on your machine.

The app's real value is the density of controls it exposes over the raw prompt. You get fine-grained sampler settings, prompt manager with drag-and-drop ordering, author's notes, world info (lorebooks) that inject context on keyword triggers, character cards (V2/V3 spec) with example dialogue and personality fields, user personas, group chats with turn management, chat branching, swipes/regenerates, and a QuickReply/STscript scripting engine for macros and slash commands. RAG-style document ingestion, vector storage, and web-search extensions ship as first-party plugins. Extras add TTS (ElevenLabs, XTTS, Silero, OpenAI TTS), image generation (SD-WebUI, ComfyUI, Novita, OpenAI, Stability), speech-to-text, translation, and expression sprites for characters.

Typical workflows: running uncensored local models against a rich character card, orchestrating multi-character group scenes, building interactive fiction with lorebook triggers, or using it as a general chat client that beats vendor UIs on prompt control. The learning curve is real — the settings panels are dense and terminology assumes familiarity with samplers, context templates, and instruct formats — but for hobbyists who want to see and edit every token that goes to the model, nothing else in the ecosystem comes close.