Overview
The Intelligent D&D Narrator is a full-stack web application that acts as an AI-powered Dungeon Master, guiding players through dynamic Dungeons & Dragons campaigns. Built with React and Next.js, it leverages the Model Context Protocol (MCP) to give the LLM access to persistent game state, enabling rich and coherent storytelling across sessions.
Motivation
Tabletop RPGs are incredible experiences, but they require a dedicated Dungeon Master who can invest significant time in world-building and session preparation. This project aims to lower that barrier by creating an AI narrator that can improvise, remember context, and maintain narrative consistency — making D&D accessible to groups that lack a dedicated DM.
Architecture
The application is built on a modern full-stack TypeScript foundation:
- Frontend: React with Next.js for server-side rendering and routing, styled with Tailwind CSS
- Backend: Next.js API routes handling game logic and LLM communication
- Database: PostgreSQL managed through Prisma ORM for persistent campaign and character data
- AI Layer: LLM integration via MCP, giving the model access to game state, character sheets, and world lore
Key Features
- Persistent Campaigns — Save and resume campaigns across sessions with full context preserved
- Character Management — Create and track characters with stats, inventory, and backstory
- Dynamic Narration — The AI adapts its storytelling based on player choices and campaign history
- MCP Integration — The model can query and update game state directly, enabling mechanically accurate encounters
How MCP Ties It Together
The Model Context Protocol is what makes this more than a simple chatbot. Through MCP, the LLM can:
- Look up character stats and inventory before narrating combat
- Track quest progress and reference past events naturally
- Enforce game rules by checking against stored mechanics
- Update the game world based on player actions
Challenges
Maintaining narrative coherence over long campaigns was the primary challenge. LLMs tend to lose context over extended conversations, so the MCP-backed retrieval system was critical for grounding responses in actual game state rather than hallucinated details.
What I Learned
Building this project taught me a great deal about integrating LLMs into stateful applications. The MCP pattern proved to be a powerful abstraction for giving models structured access to external data without sacrificing the flexibility of natural language interaction.