docs/MCP_INTEGRATION.md

MCP Integration

Texture-Craft exposes a single MCP tool: generate_texture.

flowchart TD
    A[AI Agent / IDE] --> B[MCP Client]
    B --> C[Local FastMCP server or Gradio Space MCP]
    C --> D[TextureEngine.generate_batch]
    D --> E[Generated tile variants]
    E --> F[Base64 PNG / data URI payload]
    F --> A

Example texture set

{
  "mcpServers": {
    "texture-craft": {
      "command": "C:\\path\\to\\Texture-Craft\\.venv\\Scripts\\python.exe",
      "args": ["C:\\path\\to\\Texture-Craft\\local_mcp_server.py"]
    }
  }
}

For the local stdio workflow, texture-craft-mcp runs the FastMCP server. The REST API remains available through python server.py. The Hugging Face deployment uses app.py and Gradio's built-in MCP server; it does not start a separate FastMCP service.

The tool accepts typed arguments:

• prompt
• variants (1–8)
• seed (-1 for random or an integer from 0 to 2147483647)
• steps (10–50)
• guidance_scale (3–15)

Example payload:

{
  "prompt": "golden wand with arcane sparkles",
  "variants": 2,
  "seed": 42,
  "steps": 30,
  "guidance_scale": 9.0
}

Example response:

{
  "prompt": "golden wand with arcane sparkles",
  "count": 2,
  "textures": [
    {
      "variant_index": 1,
      "seed": 42,
      "latency_sec": 1.8,
      "filename": "texture-01.png",
      "mime_type": "image/png",
      "base64_png": "...",
      "data_uri": "data:image/png;base64,..."
    }
  ]
}

The Hugging Face Space MCP endpoint is https://<owner>-<space-name>.hf.space/gradio_api/mcp/. The public tool schema is at /gradio_api/mcp/schema; deployment instructions and verification steps are in Hugging Face Space deployment.