text-generation-webui, KoboldCpp, KoboldAI Lite

The local roleplay and creative-writing community runs on a different set of tools than the chat-assistant world. text-generation-webui (oobabooga), KoboldCpp, KoboldAI Lite, and SillyTavern are the dominant FOSS UIs and engines for character chat, long-form fiction generation, and persistent-character workflows. Different audience than Open WebUI (chat assistant) — same underlying model technology.

Cross-links: llama.cpp is the engine; TabbyAPI / Aphrodite for the EXL2 roleplay path; models for picks. See also overview.

text-generation-webui (oobabooga)

★ text-generation-webui (github.com/oobabooga/text-generation-webui, AGPL-3.0) — the long-running multi-backend webui. Often called "oobabooga" or "ooba."

• Supports llama.cpp, ExLlamaV2, Transformers, AutoAWQ, AutoGPTQ — the kitchen-sink runtime selector.
• Notebook tab, chat tab, character cards, LoRA training tab.
• Extensions ecosystem (sd_api_pictures, long_term_memory, alltalk_tts).
• Running since the early Llama-1 era; the "I want all the knobs" UI.
• Honest verdict: showing its age compared to Open WebUI for chat use, but still the broadest multi-backend front-end. Power users keep it around.

KoboldCpp

★ ★ KoboldCpp (github.com/LostRuins/koboldcpp, AGPL-3.0) — single-binary fork of llama.cpp with a built-in web UI tuned for KoboldAI-style storytelling.

• ★ Single executable — .exe on Windows, single binary elsewhere; no Python, no install.
• Built-in story / chat / instruct / adventure modes.
• Compatible with OpenAI-API and the older Kobold-API.
• Image generation integration (StableUI), TTS integration.
• Excellent for the "dump on a USB stick, run on any PC" demo.
• Best-in-class single-binary UX.

KoboldAI Lite

KoboldAI Lite — minimal browser-only UI that talks to any KoboldCpp / KoboldAI / OpenAI-compat backend. Stateless, privacy-friendly, no install.

SillyTavern

★ ★ SillyTavern (github.com/SillyTavern/SillyTavern, AGPL-3.0) — the dominant front-end for character roleplay.

• Character cards (PNG with embedded card metadata).
• Persona system, group chats, world info / lorebooks.
• Connects to any backend: KoboldCpp, TabbyAPI, Aphrodite, Ollama, OpenAI, Claude, OpenRouter, Mistral.
• Massive extension ecosystem (TTS, STT, image gen, vector storage).
• The standard for serious local-roleplay setups.

Recommended creative-writing stack (May 2026)

• Engine: TabbyAPI with EXL2 quants of a 70B model on a 24GB GPU, or KoboldCpp with GGUF on any hardware.
• Frontend: SillyTavern.
• Model picks: community favourites in May 2026 are Llama-3.3-70B-Instruct, Qwen 3 72B, Magnum-v4 (Llama-tuned for prose), DeepSeek V3 distillations, and various Llama 4 finetunes.
• Sampler tuning: Min-P, dynamic temperature, smoothing — the roleplay community has spent years tuning these.

Honest framing

This cluster is a thriving FOSS community that gets passed over by mainstream "I want a chat AI" coverage. The tools are excellent at what they do. They overlap with Open WebUI but have a different center of gravity — character persistence, long-form generation, sampler control, fewer corporate-assistant guardrails.

Pick this if…

• Character roleplay: ★ ★ SillyTavern + KoboldCpp or TabbyAPI.
• Single-binary local LLM with a UI: ★ ★ KoboldCpp.
• Maximum-knob multi-backend UI: text-generation-webui.
• Long-form fiction writing: SillyTavern with a writing-tuned model + EXL2 quant.
• You want a chat assistant, not a roleplay setup: Open WebUI instead.