Lanternbait

Lanternbait is a dungeon sandbox. You don't crawl the dungeon — you build it.

Every morning adventurers walk in for loot; every evening the survivors sell,

heal and gear up in town. Whoever is still inside when the day ends is buried

in the collapse.

The heroes play by themselves. They map the dungeon as they go and remember it

between runs: the trap that bit them, the room where a friend died, the chest

worth coming back for. When two heroes meet, they decide what to do — talk,

team up, strike a deal, rob each other, or stab a partner in the back.

MODES

• Creative — a sandbox. Dig rooms, place monsters, traps, chests, loot and

notes, write your own heroes. You can even play one hero yourself,

turn-based, alongside the AI.

• Hoard — you are the dungeon. Your gold lies in the dungeon as real piles,

and it lures heroes in. Build at night, pay your monsters, raise fallen

heroes as Revenants, and keep anyone from reaching the Heart for 30 days.

THE AI IS OPTIONAL

Lanternbait plays fully without any language model: heroes then run on

hand-written tactics. Connect a model and they start thinking out loud: they

choose goals, argue in parleys, reflect on the day, read your notes

(and decide whether to trust them), name artifacts and write epitaphs.

The rules always stay in code — the model only nudges decisions.

Supported model sources:

• a local server on your machine — Ollama, LM Studio, llama.cpp (the game can

download a small recommended model through Ollama for you);

• a paid OpenAI-compatible API with your own key (OpenAI, Anthropic,

OpenRouter, Gemini, DeepSeek, Groq, Mistral…);

• Claude Code or Codex CLI on your own subscription.

Nothing is sent anywhere unless you pick an online source yourself.

Also: a chronicle of every death, artifact and betrayal, exportable as text;

custom tilesets; ASCII by default.

This is an early prototype — expect rough edges and balance swings.

Feedback in the comments is very welcome!