The Minds Behind the Mask

The steadfast Dungeon Master — the default voice at every free table. Fast, dependable, and built for structured play: it runs the rules engine, the Guardian, and the extraction of every scene into the world state. A sparse mixture-of-experts model that wakes only a sliver of itself per token, it answers with a brisk, reliable cadence and offers three selectable reasoning levels (low, medium, high) so you can trade depth for speed.

The quick-thinking mind — an open-weight sparse MoE reasoning model from DeepSeek with 16B active parameters out of 552B total, released under the permissive MIT license. It reasons before it answers and reads the whole realm in its 1M-token context window. The pragmatic pick: strong reasoning at a fraction of the weight, at a budget price.

The Cauldron can now speak with a voice that lives on your own hardware. The server stays the brain — it builds every prompt and keeps the books: dice, HP, combat and the world state are rolled and owned server-side, exactly as before. Your machine becomes the mouth: the browser asks your local Ollama to narrate and hands the text back. No API key exists for Ollama, so there is nothing to store, nothing to leak, and the DM step costs 0 turns — it isn't our electricity anymore. It isn't our content policy either: the house minds (StepFun, Qwen, DeepSeek) are subject to their providers' policies, while a model on your hardware answers only to you. Any model your Ollama can run is fully playable as the DM — uncensored ones included. Images and voice remain house engines; this is about the DM's words.

The honest limitations. The daily allowance follows how many roles you run at home: while any background role still rides our models, local play is capped at roughly 100 local messages per account per day — still about three times the free tier's play, for none of the DM-step cost. Pick 🏮 models for both the Lorekeeper and the Background and the cap lifts to roughly 300 messages a day — which is simply because our API cost per turn drops to about zero there. One dependency, honestly stated: local background roles only take effect when the DM is local too — the relay lives on that path. Periodic summaries and the logbook stay on the house either way, and voice narration and images are not local features — they keep costing house turns as before. Narration quality is exactly as good as the model you picked and as fast as your hardware: an 8B model tells a serviceable tale, a sub-8B one will lose the plot of a long campaign, and a 14B+ thinking model on a real GPU is where this stops feeling like a trade-off — the bookkeeping roles ask even more instruction-following, since their steps come back as JSON-like tasks. The prompt your Ollama receives is the same prompt our DMs get — it is deliberately not a secret; you can inspect what your DM was actually asked. And a local campaign is your own hardware's child: dice, HP and combat remain server-authoritative, so nothing fabricated locally changes the books. The whole feature is opt-in — nothing is probed until you tick Use own Ollama. 🏮 Bring Your Own Breath — the full player guide covers setup, the CORS trap behind the yellow dot, and the troubleshooting FAQ.

The voice that belongs to the table — StepFun TTS reads your DM, characters and NPCs aloud with a brisk, reliable narration. It is the everyday narrator: fast, dependable, and there for every adventure. Part of the one-time 10€ unlock — each narration costs 1 turn.

The richer narrator — Qwen’s audio mind reads the scene with more warmth and nuance, offering a male and a female narrator (longanlufeng and longanlingxin) and a softer, more expressive pacing. When the realm should linger on its words, this is the voice. Unlocks with the one-time 10€ unlock.