How to Play

A guide to the adventure — from the first step to the last dice. The short version: speak naturally, and let the world keep the numbers.

🧠 Curious how it works under the hood? — The Machinery: the DM prompt injections, the Lorekeeper's contracts, the dice engine, TTS caching and character weaving.

What Happens When You Type

Every AI game says its dice are honest. This is the order of operations — it is the whole design.

You write

Plain words, or your own voice. No syntax, no menus, no commands to learn before the story starts.

The engine rolls first

Checks, saves and attacks are rolled on the server before the DM writes a word. The narrator is handed a finished number and may not overrule it — a miss stays a miss, however well it is described. If the server can't be reached, the roll is labelled offline instead of being quietly rolled in your browser.

The world is written down

HP, items, promises, grudges, the dead — stored as data, not as scrollback. That is why a campaign still knows itself a hundred turns and one season later.

Athletics · Strength check d20 14 + 3 STR + 2 proficiency = 19 Success beat DC 15 by 4 · server roll

Core Loop

1.The DM awakens — asks 2-3 questions about your character and the world 2.You answer — the DM weaves your answers into the opening scene 3.You act — write what you do in plain words: fight, sneak, speak, explore. No syntax, no commands — intent is enough 4.The DM responds — describes consequences, plays the NPCs, opens new choices 5.The Lorekeeper writes — HP, gold, items, XP and quests update in the background, quietly 6.Repeat — the story is built from your choices, nothing is predetermined

Two Minds at the Table

The Dungeon Master narrates everything — the world, the NPCs, the drama. It thinks in stories, not spreadsheets
The Lorekeeper is the quiet scribe behind the DM — it reads every reply and turns it into real mechanics: damage applied, treasure added, map updated, journal filled
While the DM answers you, the Lorekeeper works in parallel — you can see its progress live (Lorekeeper reviewing the action… → DM weaving the tale… → Parsing mechanics… → Lorekeeper updating the world…). The story never pauses for bookkeeping
Mechanical effects arrive as a 🦉 Lorekeeper Report in the chat — that's your proof the numbers landed

The Ledger — How the World Remembers

Everything you do is written down twice: once as the story, once as the books. The story is what the DM tells you; the books are what actually happened.

Story layer The Transcript

Every turn is appended to an unchangeable record — the DM's narration, your words, the dice. It can be replayed, exported, and searched by memory. The past is never rewritten.

Book layer The Fact Register

After every reply, the extraction model reads the scene and pulls out the durable truths — NPCs, places, promises, relationships, world changes — into a structured register. Facts that matter stay near the top of the DM's memory; forgotten details quietly sink as they stop being relevant.

Each fact is deduplicated and versioned — if the world changes, the old truth is superseded, never silently edited
Facts are injected into the DM's context every turn as an authoritative block: FACT REGISTER (authoritative — never contradict)
NPCs carry their own dossier: a relation (allied, neutral, enemy or unknown) plus notes the Guardian keeps per person — what they know, what they remember about you — surfaced when you talk to them
The 📖 Codex → Facts tab shows the full register with rank — how hot each memory is right now

The Guardian — Keeping the Truth

The DM is a storyteller, and storytellers make mistakes. The Guardian is the part of the engine that owns the numbers — so a wrong sentence in the tale never becomes a wrong fact in the world.

The rule of the books: the Guardian corrects the ledger only when the Dungeon Master's own narration states it as truth in the world — it never writes a fact the DM did not speak. Truth flows from the DM's books, mechanics flow from the Guardian. That split is what hardens the realm against AI drift: a wandering sentence can never quietly rewrite the state.

Before the story Guardian Check

When you act, the Guardian first decides what the rules say — is a roll needed? What DC? Advantage? The DM narrates within those bounds, and the dice are always rolled by real code, never invented.

After the story Guardian Extraction

It then reads the DM's reply and commits the mechanics to state: damage applied, gold moved, items added, quests advanced. The story can say anything; the books change only through the Guardian.

If a drift happens — the DM narrates something that contradicts the books (a dead NPC returns, an item you sold appears) — the authoritative fact block keeps the truth visible, and the Guardian's state wins
Every contradiction is caught by the "never contradict" rule in the DM's context plus the mechanical state the Guardian owns
You can correct the books directly: /guardian my HP should be 18 — the Guardian updates the ledger, no re-narration needed
Every 50 turns a compaction pass sweeps the fact register: stale, irrelevant details are down-ranked so the memory window stays focused on what matters now

Getting Started

Sign in at the Gate — free account, 20 fresh turns every day, no email needed. The one-time 10€ unlock adds 100 turns that never expire and opens every Dungeon Master, voice and image engine
What costs a turn: every chat message is 1 turn. An image, a voice narration, character summoning or a memory search each cost 1 turn of their own — images, voice and the other Dungeon Masters need the one-time 10€ unlock first. Undo needs a donation of any amount (1€ is enough) and stays available forever once unlocked; each use costs 1 turn. Everything else — the Guardian, the ledger, the map — is free
Choose how to begin: Prepare Adventure (describe a world and/or upload .md/.pdf files — Qwen extracts characters, locations, quests and lore) or New Adventure (the DM freestyles an opening on the spot)
Create your character: pick an archetype or describe your own, then Summon Character — the AI writes the full sheet
Pick your DM model and step to the table. The DM awakens and asks its first questions

Dice — Roll or Write

When an outcome is uncertain, the DM calls for a roll and a Roll or Write card appears in the chat
Click the 🎲 dice to roll — a short ceremony, flicker and glitter, then the verdict
Or click the 🖊️ pen to describe your attempt in words instead — the chat unlocks and you narrate freely
The chat is locked while a roll is pending, so the dice stay fair and deliberate
Natural 20 = triumph, rendered in gold. Natural 1 = catastrophe, rendered in blood
You never need to type dice notation anywhere — the DM and Lorekeeper handle the math

Talking to NPCs

Type @Name in the chat to address an NPC directly — e.g. @Lyra Vane "What are you doing here?"

The NPC answers in its own voice, with its own opinions and memory of you. The DM can step in with narration at any time.

Typing @ gives you autocomplete — every NPC you've met is listed. NPCs standing close to you join The Party panel in the sidebar.

The Codex — Everything in One Place

The 📖 Codex is your tabbed view of the whole campaign — no page-hopping:

Character — full sheet: stats, HP, inventory, equipment, notes
NPCs — every person you've met, with relation and lore
Map — locations discovered so far
Journal — the logbook of your adventure
Facts — pinned truths about the world
Spells — known spells and slots

Voice, Sound & Themes

Voice of the DM — included with the one-time 10€ unlock: a ▶ play button on every DM reply reads the narration aloud (1 turn per narration), and you pick a male or female storyteller voice above the input. Voices are generated on demand and cached
Theme switcher in the top bar — five dark-fantasy palettes (Terminal, Blood, Arcane, Ember, Iron), remembered between visits
Typewriter reveal on messages, with pixel-particle bursts on every DM, Lorekeeper and player message

Models

Three roles, each with its own model — all switchable mid-adventure from Settings (⚙️):

Dungeon Master — the storyteller. Higher reasoning models think longer but write richer prose; watch the DM's reasoning bubble if you're curious what it's weighing
Lorekeeper — the mechanic extracting numbers from the narration
Background (extraction) — used for worldbuilding uploads and avatars

Open About the Models in Settings for what each model does and why thinking takes time. Your choices are saved per campaign. If the DM stops responding, switching models is the first thing to try.

Running the DM on Your Own Machine

How do I run the DM on my own machine? Install Ollama, pull a model, and pick it in Settings (⚙️) under the 🏮 group — e.g. ollama pull qwen3:8b. The server still builds the prompts and keeps the books (dice, HP and combat stay server-side); your browser asks your own Ollama for the narration and hands it back. The DM step costs 0 turns.

Same machine — works out of the box in desktop Chrome/Firefox at http://127.0.0.1:11434. If the picker shows a yellow CORS dot, Ollama must be told the site may call it: OLLAMA_ORIGINS=https://dnd.rostad.cc ollama serve
Another box on your network — run Ollama with OLLAMA_HOST=0.0.0.0 plus the same OLLAMA_ORIGINS line and paste the address (any port) into the base-URL field in Settings; it lives only in your browser and is never sent to us. One honest caveat: this page is served over HTTPS, and browsers refuse plain-HTTP calls to a LAN address from an HTTPS page (loopback to your own machine is exempt). Reaching another box therefore needs TLS on that address — on a tailnet, tailscale serve in front of Ollama does it in a minute.
Model size — a ≥14B thinking model is the comfortable pick, 8B is the practical minimum, and anything smaller will struggle to hold a long campaign — and the full-local mode below asks even more of it: the bookkeeping steps come back as JSON-like instructions, which a tiny 3B model simply fumbles. Set the context to 16k (the slider does this per request — Ollama's own default of 4k is too small)
What stays house-run — the Lorekeeper and fact extraction still run on our models in the background, so local play is capped at ~100 messages per account per day; voice narration and images cost turns as before
Whole Cauldron local — an opt-in per-campaign setting in the 🏮 row: your own Ollama also runs the Lorekeeper's roll-check, the fact extraction, the tag repair and the memory-search continuation — same prompts, your hardware (the house keeps the periodic summaries, logbook, voice and images). That mode asks for a model that follows JSON-like instructions: a 14B+ thinking model shines, a tiny 3B one fumbles the bookkeeping. Because the house pays almost nothing per turn there, the daily cap lifts from ~100 messages to ~300. Being switched on gradually

Full troubleshooting guide → — the CORS trap that keeps the dot yellow, status-dot meanings, verification commands and the FAQ.

Rules (D&D 5e)

The DM runs the table on the 5e ruleset: 1d20 + modifier vs DC, initiative and turn order in combat, advantage/disadvantage, rests, death saves
You don't need to know any of it — the DM applies the rules and explains them as they matter
Unsure about a rule? Ask the DM directly — argue your case, and the table will judge
The DM is still the final judge. The rules are a framework, not a prison

Tips

Write freely — you can attempt anything (you may try, you may not always succeed)
Everything is remembered — grudges, favors, prophecies half-fulfilled. Leave for a season; the world keeps your place
Click any portrait to see it full-size — generate a new one from there (image generation is part of the 10€ unlock; 1 turn per image).
📊 Usage in Settings shows exactly which models ran and what they cost in tokens — full transparency
📦 Export in Settings downloads your entire campaign as a ZIP — included with the 10€ unlock, and unlocked forever by any donation (1€ is enough)
🚪 Leave in Settings returns you to adventure select — your campaigns wait for you there

If Something Goes Wrong

The state looks wrong? (missing item, wrong HP) — tell the Lorekeeper directly with /guardian my HP should be 18 and it corrects the books
No reply from the DM? The chat shows the underlying error — an HTTP code like HTTP 429 (rate limit), HTTP 500 (server), or a network message
If the issue persists, change models in Settings (⚙️) — a different model often routes around the problem
Connection dropped mid-reply? The chat recovers the DM's answer from the transcript automatically
Check the 🛠️ Engine Room in Settings for live debug logs if things look wrong; /help in the chat lists the few remaining shortcuts

Back to the Gate