KoboldAI is the fully-yours end of the AI dungeon master spectrum: a model running on your own hardware, a story that never leaves your machine, and zero subscription. People absolutely run D&D campaigns this way. This guide is about doing it with clear eyes, because the software gives you a narrator and nothing else.
Terminology first: most people searching "KoboldAI" today land on KoboldCpp, the actively maintained local engine that runs GGUF models and ships the browser-based KoboldAI Lite interface. That pairing is what this guide assumes. Details are current as of August 31, 2026.
What the Kobold stack gives a D&D table
A local model. Whatever your hardware runs: no API costs, no content filter you did not choose, no internet required, complete privacy. For many players this is the entire reason to be here, and it is a good one.
Adventure mode. Lite's adventure mode structures play as actions and outcomes rather than co-written prose, which is the right grammar for a game. Instruct and story modes work too if you prefer steering the narration directly.
Memory, World Info, and Author's Note. The persistence toolkit: standing facts in memory, keyword-activated lore in World Info, tone and behavior instructions in Author's Note. The same layers, and the same limits, as any prompt-based state: it is retrieval the model reads, not rules anything enforces.
What it does not give you
Dice. There is no game dice system. A roll the model narrates is a generated number that tends to flatter you. Roll real dice and paste the result, or put SillyTavern in front of KoboldCpp and use its D&D Dice extension for committed formula rolls; our SillyTavern D&D setup guide covers that build, and SillyTavern game state covers where the numbers should live.
A character sheet and rules engine. HP, slots, conditions, initiative, legal actions: all manual, all yours, forever. The Kobold stack does not drift from your ledger any less than a hosted chatbot does; it just drifts privately.
Rules-accurate narration at small sizes. Be honest about model weight. Local models that fit consumer GPUs narrate well and referee badly: modifiers misapplied, conditions forgotten, rulings invented. The setup that works treats the model as a narrator with no mechanical authority at any size, and that goes double at 8B.
The honest configurations
- Lite alone. Adventure mode, physical dice, a paper or text-file ledger. Simplest, and entirely playable solo with discipline.
- SillyTavern + KoboldCpp. The same local model behind better game tooling: real dice via extension, chat-local variables, Quick Replies, lorebooks. This is the strongest DIY build, and our guide to running your own AI dungeon master situates it in the wider self-hosting landscape.
- Local narration, hosted game. Keep the Kobold stack for freeform and private play, and run the D&D campaign in a purpose-built game.
On that last one, the trade is the same one this whole cluster keeps arriving at. In TableForge, the game engine runs the game, and the AI DM narrates on top of it: dice, sheets, combat, and campaign state are owned by code, multiplayer for up to six works in the browser, and nobody at the table is the integration layer. You give up local inference and model choice; you stop being the referee. How the rules engine works explains what that buys, and how a TableForge campaign runs shows the table it produces.
If the machine humming under your desk is half the joy, keep it in the rotation; it does freeform brilliantly. Just do not let the campaign you care about depend on the part of the stack that was never built to protect it.