How the AI works
Celeste has a Dungeon Master built in. It can run on your own machine for nothing, it asks before it does anything it cannot undo, and it is not allowed to invent the parts that matter — the rules are computed by the engine, not improvised by a model.
4 of the 6 models cost nothing. The other two run on an API key you own — never one of ours.
What it actually does at the table
It runs the world while you sleep
Rival factions take their season without you: they act, trade, declare war, recruit, and march. You read the digest when you sit back down.
It knows your world, not just D&D
Your world entries are indexed and searched by meaning, so the advisor answers from your lore — your factions, your provinces, your history.
It resolves the maths, you keep the story
Revenue, morale, mass combat, investiture DCs — the engine computes them. The model spends its attention on narration instead of arithmetic.
It proposes, you decide
In Guided DM the AI narrates and a human Guide approves. In Assistant DM you run the game and it answers. Either way the table is yours.
Six models. Four of them free.
One setting per campaign decides which brain runs your table. Change it mid-session; the next reply uses the new one.
| Model | Runs on | Cost |
|---|---|---|
| Qwen 3.6Default The default. Nothing leaves your computer, and nothing is ever billed. | Your machine, through Ollama | Free |
| Gemma 4 · 12B Three times lighter than Qwen and just as good at the table — our own bench put it ahead on tool use. | Your machine, through Ollama | Free |
| Gemma 4 · 31B A bigger model without a bigger machine. Served through the same Ollama you already run. | Ollama cloud, on your account | Free tier |
| GLM 5.2 The heaviest of the free options, for tables that want more narration. | Ollama cloud, on your account | Free |
| Claude Sonnet 4.5 Fast and sharp. The everyday paid choice. | Anthropic, on your API key | $3 / $15 per million tokens |
| Claude Opus 4.7 The richest Dungeon Master Celeste can give you, with extended thinking. | Anthropic, on your API key | $15 / $75 per million tokens |
The two cloud models are served through the Ollama you already installed, on your own Ollama account — there is no third-party key to paste and no Celeste server in between.
Four things it will never do without your word
These stop and wait for a Dungeon Master to click. It does not matter which model you chose — the gate is in the engine, not in the prompt, so a local model and Claude are held to exactly the same line.
Destroy a bloodline
A Tighmaevril weapon can end a bloodline permanently. The engine will not do that on a model’s say-so.
Take a province
Occupy, pillage, liberate or raze — holdings are deleted and land changes hands.
Cast a realm spell
Realm magic moves the loyalty of a province, and loyalty is the whole game.
Overcharge a bloodline
Spending past the pool means Blood Overcharge — exhaustion, and at the far end a player character can die.
The parts it is not allowed to make up
A Dungeon Master that invents a rule breaks the game quietly. Celeste takes the guessable parts away from the model and hands them to the engine.
It cannot invent an identifier
Every id the AI passes to the engine is checked before anything runs. A name where an id belongs is rejected with an error that tells it to stop, not to guess again.
It cannot invent an action
Domain actions come from a closed list of the sixteen the rulebook defines. Anything outside it is refused at the executor, not merely discouraged in a prompt.
It cannot invent a price
The model proposes the intent; the engine calculates the cost from the rulebook and overwrites whatever the model claimed. Fortify costs six because chapter six says six.
It cannot quietly fall back
If your local model is down, Celeste says so. It will not silently reroute to a paid API and hand you the bill afterwards.
You always know what it costs
A new campaign starts on the free local model. Nothing is billed until you deliberately paste a key and pick a paid one.
When you do, Celeste tracks every call by model and shows you the tokens, the rate and the running total. You can set a daily ceiling in dollars: the engine checks it before it makes the call, so the limit stops the spend instead of reporting it after the fact.
See it running a kingdom
The AI earns its keep in the realm game — seasons, armies, trade and treachery, played out by factions that do not wait for you.