Overview

MythWeaver is a multimodal, agentic storytelling system built with Gradio and the Model Context Protocol (MCP). The core innovation is a custom Narrative Engine MCP server that exposes context-aware narrative generation as a reusable tool for ANY application. Users interact through text and optional map/sketch uploads. The system interprets drawings, converts them into structured scene graphs, generates styled maps, and uses the Narrative Engine MCP to dynamically generate objectives, plot hooks, and narrative opportunities.

Novel Contribution: The Narrative Engine MCP server provides domain-agnostic narrative generation tools that can be used by games, interactive fiction, writing tools, tabletop RPG assistants, and educational applications.

---

Features

Core MCP Innovation

• Narrative Engine MCP Server - Novel, reusable MCP tool for narrative generation
  • Analyzes structured context (world state, entities, relationships, scene graphs)
  • Generates narrative objectives, plot hooks, conflicts, and resolutions
  • Domain-agnostic design works for stories, games, education, writing tools
  • Exposes standardized MCP interface with multiple tools

Application Features

• Text-based adventures with user-steered or semi-autonomous narration
• Sketch / map upload for geography-driven storytelling
• Vision parsing into structured scene graphs (Gemini Vision MCP)
• High-resolution map generation (Flux/SDXL MCP)
• Long-context world memory (Memory Store MCP)
• Gradio 6 interface with mobile support

---

Architecture (Text Diagram)

User
 └── Gradio UI
       ├── Text Input
       ├── Image Upload
       └── Quest Selection
             │
             ▼
      Claude 4.5 Sonnet (Anthropic, Orchestrator)
             ├── World-memory reasoning
             ├── Narrative generation
             ├── Tool selection & routing
             └── Multi-step planning
             │
             ├── ⭐ Narrative Engine MCP (CUSTOM - Our Novel Contribution)
             │       → generate_narrative_objectives(context, genre, count)
             │       → analyze_narrative_opportunities(scene_graph, entities)
             │       → Returns structured objectives with narrative hooks
             │
             ├── Gemini Vision MCP (Google)
             │       → parse_scene(image) → Scene Graph
             │
             ├── Map Generator MCP (Flux/SDXL)
             │       → generate_map(scene_graph) → Styled Map Image
             │
             └── Memory Store MCP
                     → update_world(locations, NPCs, events)
                     → query_world() → World State
             │
             ▼
      Gradio Output Display
             ├── Story Text
             ├── Generated Maps
             └── Quest Options

Key: ⭐ = Novel MCP server built for this hackathon

---

Tech Stack

Frontend

• Gradio 6 - Chat interface with image upload and quest selection

Orchestrator Model

• Claude 4.5 Sonnet - Planning, narrative generation, tool routing

MCP Servers (4 Core Tools)

MCP Server	Purpose	Status
⭐ Narrative Engine MCP	Context-aware narrative generation	Custom - Our contribution
Gemini Vision MCP	Scene graph extraction from sketches	Existing
Map Generator MCP	High-resolution map generation	Existing/Custom
Memory Store MCP	Persistent world state	Existing

Narrative Engine MCP Server Details

Tools Exposed:

# Tool 1: generate_narrative_objectives
Input:
  - context: dict         # World state, entities, relationships
  - narrative_style: str  # Genre/tone (fantasy, sci-fi, mystery, etc.)
  - count: int           # Number of objectives (1-5)
  - difficulty: str      # easy, medium, hard

Output:
  - List of objectives with:
    - description: str
    - narrative_hook: str       # Why this matters to the story
    - success_criteria: list
    - conflicts: list           # Potential obstacles/antagonists
    - related_entities: list    # NPCs, locations involved
    - estimated_complexity: str

# Tool 2: analyze_narrative_opportunities
Input:
  - scene_graph: dict    # Structured scene data
  - character_state: dict # Character relationships, goals, conflicts

Output:
  - List of narrative opportunities:
    - opportunity_type: str  # conflict, mystery, romance, discovery, etc.
    - description: str
    - involved_entities: list
    - dramatic_potential: float

Implementation:

• Python MCP server using mcp SDK
• LLM-powered analysis (Claude for reasoning, structured outputs)
• Pydantic schemas for type safety
• Domain-agnostic design - works for any narrative context

---

Agentic Capabilities

• Multi-step narrative planning via Claude orchestration
• Autonomous tool routing based on user input
• Scene graph reasoning and validation
• World memory management and updates
• Context-aware narrative generation via Narrative Engine MCP
• Semi-autonomous storytelling flow

---

User Flow

1. User starts an adventure with text or uploads a sketch/map
2. If image uploaded:
   • Gemini Vision MCP → extracts scene graph
   • Map Generator MCP → creates styled map
3. Memory Store MCP → updates world state
4. Claude orchestrator → generates story continuation
5. Narrative Engine MCP → analyzes context and generates 1-3 dynamic objectives
6. Gradio displays: story text, map, and quest options
7. User selects a quest or continues exploration
8. Loop repeats with updated memory and context

---

MCP Tools Summary

Tool	Purpose	Type
⭐ Narrative Engine MCP	Context-aware narrative generation	Custom
Gemini Vision MCP	Sketch → scene graph	Existing
Map Generator MCP	Scene graph → map image	Existing/Custom
Memory Store MCP	Persistent world state	Existing

---

Setup (uv)

1. Install dependencies

uv sync

2. Run the app

uv run python app.py

---

Suggested Project Structure

/mythweaver
  ├── app.py                          # Main Gradio application
  ├── README.md
  ├── pyproject.toml
  ├── .gitignore
  │
  ├── /mcp_servers
  │     └── narrative_engine/         # ⭐ Custom Narrative Engine MCP Server
  │           ├── server.py           # MCP server implementation
  │           ├── schemas.py          # Pydantic models for narrative objects
  │           ├── generator.py        # LLM-powered narrative logic
  │           └── README.md           # Standalone documentation for reuse
  │
  ├── /mcp_clients                    # Client integrations for existing MCP servers
  │     ├── gemini_vision.py          # Gemini Vision MCP client
  │     ├── map_generator.py          # Map Generator MCP client
  │     └── memory_store.py           # Memory Store MCP client
  │
  ├── /prompts
  │     ├── orchestrator.md           # Claude system prompt
  │     └── narrative_generation.md   # Narrative Engine prompt
  │
  ├── /demos
  │     └── cli_demo.py               # Standalone CLI using Narrative Engine MCP
  │
  └── /assets
        └── example_maps/             # Sample sketches for testing

---

Hackathon Submission Checklist

• Custom Narrative Engine MCP Server - Novel, reusable tool with clear API
• MythWeaver Gradio app - Primary demo using Narrative Engine
• CLI demo - Second app proving reusability
• Multimodal pipeline (sketch → scene graph → map → narrative objectives)
• Autonomous agent behavior via Claude orchestration
• Standalone MCP documentation - So others can integrate it
• Demo video (1-5 minutes showing both apps)
• Social media post with #MCPBirthday hashtag
• GitHub repo submitted to MCP-1st-Birthday organization

Why This Is Novel

1. First Narrative Generation MCP Server - No existing MCP tool does context-aware story generation
2. Domain-Agnostic Design - Works for fantasy, sci-fi, mysteries, education, etc.
3. Proven Reusability - Two separate demo apps using the same MCP server
4. Standardized Interface - Clean APIs that any MCP client can call
5. Multi-Agent Orchestration - Shows MCP servers working together in complex workflow

---

Demo Flow

1. Start adventure - Text input or sketch upload
2. Vision parsing - Gemini Vision MCP extracts scene graph
3. Map generation - Map Generator MCP creates styled map
4. Story generation - Claude orchestrator writes narrative
5. Objective generation - Narrative Engine MCP analyzes context and proposes 1-3 narrative objectives
6. User interaction - Select quest or continue exploration
7. Memory update - Memory Store MCP persists world changes
8. Loop - Repeat with updated context

Expected Impact

The Narrative Engine MCP server can be adopted by:

• Game developers - Dynamic quest/objective generation for RPGs and adventure games
• Interactive fiction platforms - Branching narrative suggestions (Twine, ChoiceScript, etc.)
• Writing tools - Context-aware plot suggestions and story prompts (Scrivener, Obsidian plugins)
• Tabletop RPG tools - AI Dungeon Master assistance (Roll20, Foundry VTT)
• Educational apps - Generate story-based learning scenarios
• Creative AI apps - Any app needing narrative context understanding