Connect your own AI
In Bring Your Own Game Master, the dungeon's game master is an AI model, and it is yours. It runs on your computer (or your own Claude plan) through BYOLLM, a small free program. The game never learns which model you use, and it costs us nothing per player.
1. A model to run the dungeon
We have tested the game with these two:
Gemma 4 26B (A4B), on llama.cpp
Free, and runs entirely on your machine. It needs a graphics card with about 16 GB of memory or more, or a Mac with 32 GB of memory. In our tests a GM turn took about 11 seconds.
Claude, through the claude command-line tool
Needs a Claude Pro or Max plan, with Claude Code installed and signed in. Nothing else to download. A GM turn took about 20 seconds, and it uses your plan's limits.
Other models can work too. Smaller ones break the game's rules more often; while you play, "GM health" under the game says how yours is doing.
2. BYOLLM on your computer
- Open Terminal (Mac) or PowerShell (Windows) and run
node --version. If that fails, or shows a version below 22.14, install the LTS version of Node.js from nodejs.org, then open a new window. - Run
npm install -g byollm@latest - Run
byollm setup. It asks a few questions: it finds the model you have (llama.cpp, Ollama, MLX, or your Claude sign-in) and pairs with byollm.cloud. When it prints a pairing code, approve it at dashboard.byollm.cloud/devices and check the fingerprint matches. - Run
byollm status. If BYOLLM isn't running, runbyollm start.
More help: docs.byollm.cloud.
3. Connect this game
Press the button and say yes on byollm.cloud. When it asks which model should be this game's "Game master", pick the one from step 1.
Checking this browser's connection...
4. A test ruling
Before you play, send your model one GM turn: the guard room, and "I shout at the goblins to clear out". It usually takes 10 to 30 seconds; the first can take longer while your model loads. We wait up to two minutes. If nothing answers by then, BYOLLM isn't running, or isn't connected to this game.