DeepMTG: Building 'Magic: The Gathering' decks using LLM chains.

See this blog-post for more information

Installation

For development usage, we use poetry to handle dependency installation. Poetry can be installed via, e.g.

curl -sSL https://install.python-poetry.org | python3 -
poetry self update

and ensuring that poetry is available in your $PATH

Install Ollama, e.g. curl -fsSL https://ollama.com/install.sh | sh

Pull Ollama models, e.g.:

ollama pull MFDoom/deepseek-r1-tool-calling:8b
ollama pull phi4:latest

Install the dependencies:

poetry install
poetry self add poetry-plugin-export
poetry config warnings.export false
poetry run pre-commit install

Data

Download these and place them in the data directory:

• rules [https://media.wizards.com/2024/downloads/MagicCompRules%2020241108.pdf]

Download the desrired .json files and place them in the data/cards directory:

• cards [https://mtgjson.com/downloads/all-sets/]

Usage

To run the program, use the following command:

poetry run python scripts/build_deck.py "prompt for you deck"

The deck will be saved to the current directory, with a generated name.

Extra options include:

• --deck-llm-name to specify the LLM used for deck generation (default: MFDoom/deepseek-r1-tool-calling:8b)
• --summary-llm-name to specify the LLM used for card-summary generation (default: phi4:latest)

N.B. The first run will take a while as the cards will be summarised and embedded. After the first run, the embeddings will be cached and the process will be much faster.

N.B. Deck generation may be slow, depending the LLM used.

TO DO:

• Improve initial manabase generation
• Add card-type filtering to searches
• Add post-constuction deck refinement
• Add deck-performance prediction