GENTLE

Game Engine Navigation Tool for Learning & Exploration

An MCP server providing instant semantic search across game development documentation. No more copy-pasting docs into your LLM - GENTLE gives your AI assistant direct access to Unreal Engine and PyQt6 documentation.

Overview

%%{init: {'theme': 'base', 'themeVariables': {'primaryColor': '#1e40af', 'primaryTextColor': '#ffffff', 'primaryBorderColor': '#3b82f6', 'lineColor': '#3b82f6', 'secondaryColor': '#1e3a5f', 'tertiaryColor': '#0f172a', 'background': '#000000', 'subGraphLabelBgColor': '#1e40af', 'clusterBkg': '#0a0a0a', 'clusterBorder': '#3b82f6'}}}%%
graph LR
    subgraph "Your AI Assistant"
        LLM[Claude / GPT / etc]
    end

    subgraph "GENTLE MCP Server"
        MCP[MCP Protocol Handler]
        EMB[Embedder<br/>MiniLM-L6-v2]
        VS[Vector Store<br/>LanceDB]
    end

    LLM <-->|JSON-RPC| MCP
    MCP --> EMB
    EMB --> VS
    VS -->|Top K Results| MCP

Features

• Semantic Search - Find relevant docs using natural language, not keyword matching
• Fully Local - No API keys, no internet required at runtime
• Fast - Sub-second queries against 68k+ documentation entries
• Multi-Source - Unreal Python API, Console Commands, PyQt6 Reference & Tutorials

Documentation Sources

| Source | Entries | Description | |--------|--------:|-------------| | unreal-python | 60,563 | Classes, methods, properties, enums from UE5 Python API | | unreal-console | 8,055 | Console commands (stat, r., t.) and variables | | pyqt-reference | 36 | Official PyQt6 reference (signals, slots, properties) | | pyqt-tutorials | 53 | Practical PyQt6 tutorials with code examples | | Total | 68,707 | |

Tools

| Tool | Description | |------|-------------| | search_unreal_python | Search Unreal Engine Python API documentation | | search_unreal_console | Search console commands and variables | | search_pyqt_reference | Search PyQt6 official reference docs | | search_pyqt_tutorials | Search PyQt6 practical tutorials | | search_all | Search across all documentation sources | | list_sources | List available sources with entry counts |

Installation

No build steps required - embeddings are pre-built and included in the package.

Claude Desktop

Add to your config file:

• Linux: ~/.config/Claude/claude_desktop_config.json
• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "gentle": {
      "command": "npx",
      "args": ["gentle-mcp"]
    }
  }
}

Cursor

Add to .cursor/mcp.json in your project root (or global settings):

{
  "mcpServers": {
    "gentle": {
      "command": "npx",
      "args": ["gentle-mcp"]
    }
  }
}

Cline (VS Code Extension)

Open VS Code Settings → Cline → MCP Servers, then add:

{
  "gentle": {
    "command": "npx",
    "args": ["gentle-mcp"]
  }
}

Windsurf

Add to ~/.windsurf/mcp_config.json:

{
  "mcpServers": {
    "gentle": {
      "command": "npx",
      "args": ["gentle-mcp"]
    }
  }
}

Claude Code (CLI)

Add to ~/.claude/settings.json:

{
  "mcpServers": {
    "gentle": {
      "command": "npx",
      "args": ["gentle-mcp"]
    }
  }
}

Global Installation (Optional)

If you prefer a global install instead of npx:

npm install -g gentle-mcp

Then use "command": "gentle-mcp" with no args in your config.

Usage Examples

Once configured, ask your AI assistant questions like:

| Query | Best Tool | |-------|-----------| | "How do I spawn an actor at a location?" | search_unreal_python | | "What console command shows FPS?" | search_unreal_console | | "Create a button with click handler" | search_pyqt_tutorials | | "What signals does QLineEdit emit?" | search_pyqt_reference |

Architecture

Search Flow

%%{init: {'theme': 'base', 'themeVariables': {'primaryColor': '#1e40af', 'primaryTextColor': '#ffffff', 'primaryBorderColor': '#3b82f6', 'lineColor': '#3b82f6', 'actorBkg': '#1e40af', 'actorTextColor': '#ffffff', 'actorBorder': '#3b82f6', 'signalColor': '#3b82f6', 'signalTextColor': '#ffffff', 'background': '#000000', 'sequenceNumberColor': '#ffffff'}}}%%
sequenceDiagram
    participant User
    participant LLM as AI Assistant
    participant MCP as GENTLE
    participant EMB as Embedder
    participant DB as LanceDB

    User->>LLM: "How do I spawn an actor?"
    LLM->>MCP: search_unreal_python({query: "spawn actor"})
    MCP->>EMB: embedText("spawn actor")
    EMB-->>MCP: [0.23, -0.41, ...] (384 dims)
    MCP->>DB: vectorSearch(embedding, limit=5)
    DB-->>MCP: Top 5 chunks by cosine similarity
    MCP-->>LLM: {results: [...], totalResults: 5}
    LLM->>User: "Use Actor.spawn() method..."

Data Pipeline

%%{init: {'theme': 'base', 'themeVariables': {'primaryColor': '#1e40af', 'primaryTextColor': '#ffffff', 'primaryBorderColor': '#3b82f6', 'lineColor': '#3b82f6', 'secondaryColor': '#1e3a5f', 'tertiaryColor': '#0f172a', 'background': '#000000', 'subGraphLabelBgColor': '#1e40af', 'clusterBkg': '#0a0a0a', 'clusterBorder': '#3b82f6'}}}%%
flowchart TD
    subgraph Sources
        UP[Unreal Python API<br/>unreal.py]
        UC[Console Commands<br/>console-commands.md]
        PR[PyQt6 Reference<br/>reference.md]
        PT[PyQt6 Tutorials<br/>tutorials.md]
    end

    subgraph Parsing
        P1[parse-unreal-python.ts]
        P2[parse-unreal-console.ts]
        P3[parse-pyqt.ts]
    end

    subgraph "Embedding Pipeline"
        CHK[DocChunks<br/>68,707 entries]
        EMB[MiniLM-L6-v2<br/>384 dimensions]
        VEC[Vectors + Metadata]
    end

    subgraph Storage
        LDB[(LanceDB<br/>~100MB)]
    end

    UP --> P1
    UC --> P2
    PR --> P3
    PT --> P3

    P1 --> CHK
    P2 --> CHK
    P3 --> CHK

    CHK --> EMB
    EMB --> VEC
    VEC --> LDB

Chunk Structure

Each documentation entry is stored as a chunk:

interface DocChunk {
  id: string;           // "unreal-python:class:Actor"
  source: DocSource;    // "unreal-python" | "unreal-console" | ...
  type: DocType;        // "class" | "method" | "property" | ...
  name: string;         // "Actor"
  parentName?: string;  // "EditorUtilityLibrary"
  content: string;      // Full documentation text
  version?: string;     // "5.4"
  embedding: number[];  // 384-dim vector
}

Development

Setup

git clone https://github.com/yourusername/gentle-mcp
cd gentle-mcp
npm install

Scripts

| Script | Description | |--------|-------------| | npm run dev | Run development server with hot reload | | npm run build | Compile TypeScript to dist/ | | npm run fetch | Download source documentation | | npm run parse | Parse sources into chunks | | npm run embed | Generate embeddings and populate LanceDB | | npm run setup | Run fetch + parse + embed | | npm run check | Run typecheck and linting |

Project Structure

gentle-mcp/
├── src/
│   ├── index.ts          # MCP server entry point
│   ├── embedder.ts       # MiniLM embedding wrapper
│   ├── db/
│   │   └── vector-store.ts   # LanceDB interface
│   ├── tools/
│   │   └── search.ts     # Search tool implementations
│   └── types/
│       └── index.ts      # TypeScript type definitions
├── scripts/
│   ├── fetch-sources.ts  # Download documentation
│   ├── parse-all.ts      # Parse all sources
│   ├── embed-all.ts      # Generate embeddings
│   └── parsers/          # Source-specific parsers
├── sources/              # Raw documentation files
├── data/
│   ├── chunks/           # Parsed JSON chunks
│   └── lancedb/          # Vector database
└── dist/                 # Compiled output

Requirements

• Node.js 24+
• ~100MB disk space (embeddings + vector database)

How It Works

1. Parsing: Documentation files are parsed into semantic chunks (classes, methods, etc.)
2. Embedding: Each chunk is embedded using MiniLM-L6-v2 (local, ~30MB model)
3. Storage: Embeddings + metadata stored in LanceDB (local, columnar, fast)
4. Search: User queries are embedded and matched via cosine similarity
5. Response: Top-K most relevant chunks returned to the LLM

Everything runs locally. No external API calls at runtime.

License

MIT