Skip to content

MTG Judgebot

Magic rules questions answered with citations from the Comprehensive Rules, Scryfall rulings or the card's Oracle text, each checked against its source before the answer is shown.

Ambiguous names are asked about

Nicknames like “bob” and “goyf” resolve to cards through a fixed order of lookups. A name that could mean several cards becomes a “did you mean…?” prompt instead of a guess.

Citations are validated

A quote has to appear verbatim in the rule, ruling or Oracle text it names. A citation that fails is rejected and the model is asked again, and an answer with no citation is not shown.

Spend is capped

Every model call is metered against a spend cap, per process or per day or month. The worst-case cost is reserved before each call. The web app is also rate limited per IP.

Models are configuration

Anthropic direct, through a proxy, on AWS or Vertex, or any OpenAI-compatible server including a local Ollama. All of it is set in judge.toml, with no code changes.

Each community runs its own judgebot, so the spend cap, the judge role and the Discord application are yours. The web app and the command line work before you have a bot token. Read what the judge is, or go straight to requirements and first run.