rpg-dice-mcp

Go MCP server for tabletop RPG dice rolls. The One Ring 2e first; D&D / Traveller / Call of Cthulhu / etc. as plug-in packages later.

Why this exists

LLMs are famously bad at generating random numbers — they bias toward training-data distribution shape (d20 rolls cluster 14-17; "pick 1-10" picks 7 ~30% of the time). Worse, they often elide mechanical resolution entirely in narrative-heavy contexts, narrating outcomes without rolling.

Empirical evidence from a past One Ring campaign on this homelab (qwen3.6:35b, chat 491ccb24-...): 56 assistant turns, only 10 explicitly framed the Feat die, vs 32 combat-narration moments. A ~3:1 ratio of combat scenes to actual mechanical resolution.

This MCP fixes both halves of the problem:

1. Real RNG — crypto/rand-seeded math/rand, χ² uniformity verified by test for d6 / d12 / d20.
2. Forcing function — when the model HAS to call roll_tor_check, the tool call appears in the response's tool_calls field; the mechanics happened, verifiably. No more skipped rolls.

See dvystrcil/homelab#201 for the full scoping (TDD discipline, 11 acceptance criteria, three-repo split rationale).

Tools exposed

Tool	Args	Returns
roll	spec: "NdM±K"	rolls, modifier, total
roll_tor_check	skill_rating, target_number, weariness?, miserable?	feat_die, gandalf_rune, eye_of_sauron, success_dice, effective_dice, total, succeeds, margin, miserable_eye
roll_tor_combat	attacker_skill, defender_tn, weariness?, miserable?	feat_die, gandalf_rune, eye_of_sauron, success_dice, total, hits, margin, miserable_eye

More TOR-side tools (roll_eye_awareness, roll_shadow_test) and other systems land in follow-up PRs. Each new system is its own internal/<system>/ package composing on internal/dice.

Mechanics encoded (TOR 2e)

• Feat die: d12. Face 11 = Eye of Sauron (counts as 0 in totals). Face 12 = Gandalf rune (automatic success regardless of TN).
• Success dice: d6 × skill_rating. Sum contributes to total.
• Weariness: Success die rolls 1-3 are treated as 0.
• Miserable: Eye result sets miserable_eye for the caller's Hope/Shadow bookkeeping.
• Total: feat_die_contribution + sum(effective_success_dice). Compared to TN; >= TN = succeeds. Gandalf auto-succeeds regardless.

Resolution rules live in internal/tor/tor.go; rules tests are in internal/tor/tor_test.go.

Run locally

# Build
go build -o /tmp/rpg-dice-mcp ./cmd/rpg-dice-mcp

# Stdio (works with Claude Code's MCP stdio bridge)
/tmp/rpg-dice-mcp

# HTTP streamable transport
/tmp/rpg-dice-mcp -http :8080
# health probe
curl -s http://localhost:8080/healthz

# Test
go test -race -cover ./...

Trust audit

Run bin/mcp-trust-audit.py --mcp rpg-dice-mcp after deploy per the homelab MCP trust model. Stdio bridge + cluster-internal-only Service.

Repo layout

rpg-dice-mcp-docker/        ← this repo (build: Go + Dockerfile + CI)
rpg-dice-mcp/               ← deploy repo (kustomize manifests)
argocd-projects/rpg-dice-mcp/  ← Argo Application entry

Three-repo split per dvystrcil/homelab project_mcp_repo_split_pattern.

License

MIT — see LICENSE.