NanoBanana MCP Server

MCP server for AI image generation using NanoBanana Pro 2. Supports Pro (maximum quality) and Flash (fast generation) models with default 4K resolution.

Inspired by the nano-banana naming convention used across the MCP community. This is an independent implementation.

Quick Install

pip install nanobanana-imagen-mcp
# or
uvx nanobanana-imagen-mcp --help

Claude Desktop (claude_desktop_config.json):

{
  "mcpServers": {
    "nanobanana": {
      "command": "uvx",
      "args": ["nanobanana-imagen-mcp"],
      "env": {
        "GEMINI_API_KEY": "your_key_here"
      }
    }
  }
}

What It Does

Generates AI images using NanoBanana Pro 2 models. Supports text-to-image generation, image editing with reference images, file uploads, and server maintenance. Produces high-quality images at up to 4K resolution.

Tools (4)

Models

Configuration

Storage

Generated image metadata is tracked in a SQLite database.

Opt out of persistent storage by setting NANOBANANA_DB_PATH=:memory:. This uses an in-memory database that is discarded when the server exits — useful for ephemeral or read-only environments. Trade-off: image history is lost on restart; the show_output_stats tool will show an empty database each session.

The database and output directories are created lazily on first image generation — starting the server with no GEMINI_API_KEY set (e.g. nanobanana-imagen-mcp --help) does not create any files or directories.

Claude Desktop Configuration

From pip (recommended)

pip install nanobanana-imagen-mcp

{
  "mcpServers": {
    "nanobanana": {
      "command": "uvx",
      "args": ["nanobanana-imagen-mcp"],
      "env": {
        "GEMINI_API_KEY": "your_key_here"
      }
    }
  }
}

From this repository

{
  "mcpServers": {
    "nanobanana": {
      "command": "python",
      "args": ["-m", "servers.nanobanana.nanobanana_mcp_server.server"],
      "env": {
        "GEMINI_API_KEY": "your_key_here"
      },
      "cwd": "path/to/gemini-media-mcp"
    }
  }
}

Features

• Default 4K resolution for maximum detail
• Pro model for complex, photorealistic scenes
• Flash model for fast iterations and simple subjects
• Reference image support for consistent multi-angle shots
• Negative prompt support for quality control
• Thinking level control for complex multi-element scenes
• Google Search grounding for real-world subjects

License

MIT

Credits

Inspired by the nano-banana naming convention used across the MCP community. This is an independent implementation.