mnehmos.rpg.mcp: Agentic Embodied Simulation Kernel

A rules-enforced RPG backend that turns any LLM into a game master who can't cheat.

What Is This? (Start Here)

You are the player. The AI is the dungeon master.

You talk to an AI (Claude, GPT, etc.) in natural language. You say things like "I attack the goblin" or "I search the room for traps." The AI narrates what happens and describes the world.

The difference from pure AI storytelling: This engine enforces the rules. When you attack, it actually rolls dice, checks armor class, calculates damage, and updates HP in a real database. The AI can't just decide you hit or miss—the math happens, and both you and the AI see the result.

What can you actually do?

• Explore procedurally generated worlds with 28+ biome types

• Fight enemies using D&D 5e-style combat (initiative, AC, damage rolls, death saves)

• Cast spells with real slot tracking—if you're out of slots, you can't cast

• Manage inventory with equipment slots, weight, and item properties

• Complete quests with tracked objectives and rewards

• Interact with NPCs who remember your conversations across sessions

• Everything persists—close the game, come back tomorrow, your character is exactly where you left them

Who is this for?

• Solo RPG players who want AI-driven adventures with mechanical integrity

• People frustrated with AI RPGs that fall apart when you ask "wait, how much HP do I have?"

• Developers building AI game integrations who need a reference implementation

How do I play?

1. Install the MCP server (see Installation below)

2. Connect it to Claude Desktop (or any MCP-compatible client)

3. Tell the AI: "Let's start a new game. Create a character for me."

4. Play naturally—the AI handles narration, the engine handles mechanics

v1.0 Release (January 2026)

85% Tool Reduction: 195 → 28 Consolidated Tools

This release consolidates 195 individual tools into 28 action-based tools using:

• Action enums - Each tool handles multiple operations via an action parameter

• Fuzzy matching - Typo-tolerant action matching with suggestions

• Guiding errors - Clear feedback when actions don't match

Before: create_character, get_character, update_character, delete_character, list_characters... After: character_manage with actions: create, get, update, delete, list, search

Key Metrics

Meta-Tools for Discovery

Two special tools help LLMs discover and use the consolidated tools:

Architecture Philosophy

This engine implements the Event-Driven Agentic AI Architecture:

The Embodiment Model

Key invariant: LLMs propose intentions. The engine validates and executes. LLMs never directly mutate world state.

Consolidated Tools Reference (28 Tools)

Character & Party

Combat System

Inventory & Economy

World & Spatial

Quests & NPCs

Magic & Rest

Utility & Meta

Installation

Option 1: npm (Recommended)

Option 2: Standalone Binaries

Download pre-built binaries from Releases:

Option 3: From Source

MCP Client Configuration

Claude Desktop / MCP Clients:

Using Binary:

Core Systems

Combat & Encounters

• Initiative tracking with advantage/disadvantage

• Spatial combat with grid positioning and collision

• Opportunity attacks with reaction economy

• Death saving throws (D&D 5e rules)

• Damage resistance/vulnerability/immunity

• Legendary creatures with lair actions and legendary resistances

• Encounter presets - Pre-balanced encounters by party level

Magic System

• 15+ SRD spells (Magic Missile, Fireball, Cure Wounds, etc.)

• Spell slot tracking with class-based progression

• Warlock pact magic with short rest recovery

• Concentration tracking

• Anti-hallucination validation - LLMs cannot cast spells they don't know

Theft & Economy

• Stolen item tracking with heat levels (burning → cold)

• Witness recording for theft detection

• Fence NPCs with buy rates and heat capacity

• Item recognition - original owners detect their stolen goods

NPC Memory

• Relationship tracking (familiarity + disposition)

• Conversation memory with importance levels

• Context injection for LLM prompts

• Interaction history across sessions

Improvisation Engine

• Rule of Cool stunts - "I kick the brazier into the zombies"

• Custom effects - Divine boons, curses, transformations

• Arcane synthesis - Dynamic spell creation with wild surge risk

Project Structure

Design Principles

1. LLMs propose, never execute The brain suggests; the nervous system validates.

2. All action is tool-mediated No direct world mutation. Every change flows through MCP tools.

3. Validation precedes observation Act → Validate → Observe. The reflex arc pattern.

4. Deterministic outcomes Same inputs → same outputs. Always reproducible.

5. Schema-driven everything Zod validates all data at boundaries. Type safety end-to-end.

6. Anti-hallucination by design LLMs cannot cast spells they don't know or claim damage they didn't roll.

7. Token efficiency 28 consolidated tools with action routing reduce context overhead by 85%.

8. Guiding errors Invalid actions return suggestions, not just failures.

Test Coverage

Contributing

Contributions welcome! Please:

1. Fork the repo

2. Create a feature branch (git checkout -b feature/your-feature)

3. Write tests for new functionality

4. Follow existing code style (TypeScript + Zod + tests)

5. Submit a pull request

Roadmap

• Full spellcasting system with validation

• Theft and fence economy

• Corpse and loot mechanics

• NPC memory and relationships

• Improvisation engine

• Tool consolidation (195 → 28)

• Fuzzy action matching

• Preset systems (creatures, encounters, locations)

• WebSocket real-time subscriptions

• Dialogue tree system

• Cover mechanics in combat

• Quest chains with prerequisites

License

ISC — Use freely, attribution appreciated.

Related

• Model Context Protocol Specification

• Quest Keeper AI — Desktop AI dungeon master using this engine

Documentation

• CLAUDE.md - Development instructions

• docs/WHITE_PAPER.md - Design philosophy and architecture

• docs/LLMSpatialGuide.md - LLM spatial navigation guide