Nano Banana Pro MCP Server

4K Image Generation with Gemini 3 Pro Image

MCP server enabling Claude to generate and edit high-quality 4K images using Google's Gemini 3 Pro Image model (gemini-3-pro-image-preview). Features advanced text rendering, "thinking" process for complex prompts, and iterative editing capabilities.

Features

• 4K Output Resolution: Generate images up to 4096x4096 pixels
• Advanced Text Rendering: Superior text integration in images
• Thinking Process: Model reasons through complex prompts for better results
• Iterative Editing: Continue editing previous images without re-uploading
• State Management: Automatic tracking of last generated image

Tools Provided

1. nb_pro_generate_image

Generate a NEW 4K image from text prompt.

Parameters:

• prompt (string, required): Text description of the image to create

Example:

{
  "prompt": "A futuristic cityscape at sunset with flying cars and neon signs"
}

2. nb_pro_edit_image

Edit a SPECIFIC existing image file.

Parameters:

• imagePath (string, required): Full file path to the image to edit
• prompt (string, required): Description of modifications to make

Example:

{
  "imagePath": "/path/to/image.png",
  "prompt": "change the sky to purple and add mountains in background"
}

3. nb_pro_continue_editing

Continue editing the LAST generated/edited image.

Parameters:

• prompt (string, required): Description of modifications to make

Example:

{
  "prompt": "change the hat color to red and add flowers"
}

4. nb_pro_get_last_image

Get information about the last generated/edited image.

Parameters: None

Returns:

• imagePath: Path to last image
• lastPrompt: Prompt used to generate/edit
• timestamp: When image was created
• fileSize: File size in KB

Installation

Prerequisites

• Node.js 18+
• Gemini API key from Google AI Studio

Step 1: Install Dependencies

cd /Users/adriencastelain/Documents/ClaudeWorkspace/mcp-servers/nano-banana-pro
npm install

Step 2: Build TypeScript

npm run build

This compiles TypeScript to JavaScript in dist/ directory.

Step 3: Add to MCP Configuration

Add this entry to your mcp.json (or claude_desktop_config.json):

{
  "mcpServers": {
    "nano-banana-pro": {
      "command": "node",
      "args": [
        "/Users/adriencastelain/Documents/ClaudeWorkspace/mcp-servers/nano-banana-pro/dist/index.js"
      ],
      "env": {
        "GEMINI_API_KEY": "${GEMINI_API_KEY}"
      }
    }
  }
}

Note: Uses same GEMINI_API_KEY as regular nano-banana MCP server.

Step 4: Reload VSCode Window

1. Cmd+Shift+P
2. Select "Developer: Reload Window"
3. Wait 10-15 seconds for MCP to initialize

Step 5: Verify Installation

In Claude Code chat, run:

Can you list the available nano banana pro tools?

You should see 4 tools:

• nb_pro_generate_image
• nb_pro_edit_image
• nb_pro_continue_editing
• nb_pro_get_last_image

Usage Examples

Generate New Image

Generate a 4K image of a minimalist tech startup logo with abstract geometric shapes in blue and white

Edit Specific Image

Edit /path/to/image.png - change the background to dark gradient and add lens flare

Continue Editing Last Image

Now add text "INNOVATION" in bold sans-serif font centered at top

Check Last Image

What was the last image I generated with nano banana pro?

Output Directory

All generated images are saved to:

/Users/adriencastelain/Documents/ClaudeWorkspace/generated_imgs/

Filename Format:

• Generated: generated-YYYYMMDD-HHMMSS-abc123.png
• Edited: edited-YYYYMMDD-HHMMSS-abc123.png

State File: .nano-banana-pro-state.json stores last image info for continue_editing tool.

Model Details

Model: gemini-3-pro-image-preview

Capabilities:

• 4K resolution output (up to 4096x4096)
• Advanced text rendering in images
• "Thinking" process for complex prompts
• Higher quality than Gemini 2.5 Flash
• Better handling of detailed scenes

API Endpoint: Google Generative AI SDK Authentication: GEMINI_API_KEY environment variable

Comparison: Nano Banana vs Nano Banana Pro

| Feature | Nano Banana | Nano Banana Pro | |---------|-------------|-----------------| | Model | gemini-2.0-flash-exp | gemini-3-pro-image-preview | | Max Resolution | 2048x2048 | 4096x4096 | | Text Rendering | Good | Advanced | | Complex Prompts | Standard | Thinking Process | | Speed | Faster | Slower (higher quality) | | Use Case | Quick iterations | Final high-quality output |

Troubleshooting

"GEMINI_API_KEY environment variable not set"

Solution: Ensure your mcp.json has GEMINI_API_KEY in the env section and the key is set in your shell environment.

"No image data in response"

Solution: Model may not support the exact prompt. Try simplifying or rephrasing the prompt.

"Image file not found"

Solution: Use absolute file paths with nb_pro_edit_image. Or use nb_pro_continue_editing to edit the last generated image.

Build errors

Solution:

rm -rf node_modules package-lock.json
npm install
npm run build

Development

Watch Mode (auto-rebuild on save):

npm run watch

Manual Test:

export GEMINI_API_KEY="your-key-here"
npm test

Technical Architecture

Transport: stdio (standard input/output) Language: TypeScript (compiled to Node.js) Dependencies:

• @modelcontextprotocol/sdk - MCP protocol implementation
• @google/generative-ai - Gemini API client

State Management:

• Last image tracked in .nano-banana-pro-state.json
• Base64 encoding for in-memory image data
• Persistent across MCP sessions

Security Notes

• API key stored in environment variable (not in code)
• Images saved to local directory only
• No external hosting or sharing
• State file contains base64 image data (can be large)

Performance

Generation Time:

• Simple prompts: 10-20 seconds
• Complex prompts: 30-60 seconds (thinking process)

File Sizes:

• 4K images: 500KB - 2MB per image
• State file: up to 10MB (contains last image base64)

License

MIT License - Atlas (MCP Builder AI) - Boald

Version History

• v1.0.0 (2025-12-02): Initial release with 4 tools (generate, edit, continue_editing, get_last_image)

Support

For issues or feature requests, contact Atlas (MCP Builder AI) or Phoenix (DevOps).

Status: ✅ SANDBOX - Ready for testing (not in production yet)