_{Logo generated using nano-banana-mcp 🍌}

Nano-Banana MCP Server v2 🍌

An enhanced Model Context Protocol (MCP) server that provides AI image generation and editing capabilities using Google's Gemini Multimodal Image APIs (gemini-3.1-flash-image / gemini-3-pro-image).

This is a v2 fork of the original nano-banana-mcp server, updated to support modern Gemini models, custom model configuration, and direct-from-GitHub installation.

🎨 Sample Output (Imagen 4)

Here is a preview image generated using the generate_imagen tool with the default imagen-4.0-generate-001 model:

Sample Imagen 4 Output

For full visual verification and diagnostic logs for this version tag, see the TEST_REPORT.md.

✨ Features

🎨 Generate Images: Create new images from text descriptions.
✏️ Edit Images: Modify existing images using text prompts and optional reference images.
🔄 Iterative Editing: Refine the last generated or edited image sequentially.
🧠 Dynamic Model Selection: Specify which model to use via tool parameters, environment variables, or rely on a smart modern fallback.
🚀 Zero-Publish Install: Install directly from your GitHub repository using standard Git URLs.
📁 Cross-Platform Auto-Saving: Automatically saves generated images locally under platform-appropriate directories.

🛠️ Supported Gemini Models

By default, the server uses gemini-3.1-flash-image, which replaces the deprecated gemini-2.5-flash-image-preview.

You can configure or specify:

gemini-3.1-flash-image: Standard efficiency model optimized for speed and high-volume generation.
gemini-3-pro-image: High-fidelity creative model optimized for highly contextual native image creation.

🔑 Configuration & Environment Variables

The server checks configuration in the following priority:

Tool Arguments: Pass model explicitly inside tool calls (highest priority).
Environment Variables:
- GEMINI_API_KEY: Your Gemini developer token from Google AI Studio.
- GEMINI_IMAGE_MODEL: Set a default model server-wide (e.g., gemini-3-pro-image).
Global Configuration: ~/.nano-banana-config.json generated globally via the configure_gemini_token tool (with fallback/auto-migration for existing local files).

🔄 Upgrading & Key Migration from v1

If you are upgrading from the original nano-banana-mcp v1 server, your key migration is handled seamlessly:

Automatic Detection: The v2 server checks the current directory for any existing .nano-banana-config.json files from v1.
Global Persistence: If a local key is loaded and no global config exists, it automatically migrates the configuration globally to ~/.nano-banana-config.json.
Directory Independence: After running the server once in your old workspace, you can safely delete the local .nano-banana-config.json file and use the tools from any workspace.

🔑 Getting Your API Key & Google AI Studio Limits

How to Get Your API Key

Go to Google AI Studio.
Click on "Create API Key" at the top left.
Select an existing Google Cloud project or create a new one, and copy your API key.

Rate Limits & Pricing Plans (as of 2026)

Paid/Pay-As-You-Go Plan Required: Native image generation and editing using the Gemini 3 image models (gemini-3.1-flash-image and gemini-3-pro-image) are premium capabilities. Google AI Studio generally requires a Pay-As-You-Go plan for image-based generation models; standard free tiers may restrict these capabilities.
Token Billing: Image requests are billed on a pay-per-token basis, consuming approximately 1,290 tokens per generated image.
Rate Limits: Enforced at the project level. Exceeding your Requests Per Minute (RPM) or Tokens Per Minute (TPM) quotas will result in a 429: Resource Exhausted error. You can monitor and adjust your limits in the Google AI Studio console settings.

🚀 Installation & Client Integration

Method A: Build & Run Locally (Recommended for Development)

First, compile the Go binary inside the project directory:

go build -o nano-banana-mcpv2 main.go

To verify that your API key is configured correctly and image generation is functional over stdio, you can run the test script:

export GEMINI_API_KEY="your-api-key-here"
./scripts/test_generation.sh

Then add this to your MCP settings file (e.g., Cursor, Claude Desktop, or Claude Code config):

{
  "mcpServers": {
    "nano-banana-mcpv2": {
      "command": "/Users/house/Documents/gitlab/nano-banana-mcpv2/nano-banana-mcpv2",
      "env": {
        "GEMINI_API_KEY": "your-gemini-api-key-here",
        "GEMINI_IMAGE_MODEL": "gemini-3.1-flash-image"
      }
    }
  }
}

Method B: Download Pre-compiled Binary

You can download the pre-compiled binary for your system (macOS ARM64/AMD64, Linux AMD64, or Windows) directly from the GitHub Releases page.

Once downloaded, make it executable (chmod +x nano-banana-mcpv2) and add it to your path or reference it directly:

{
  "mcpServers": {
    "nano-banana-mcpv2": {
      "command": "/path/to/downloaded/nano-banana-mcpv2",
      "env": {
        "GEMINI_API_KEY": "your-gemini-api-key-here"
      }
    }
  }
}

Method C: Run via Docker

Build the minimal Docker image:

docker build -t nano-banana-mcpv2 .

Then configure your MCP client to run the server inside the Docker container:

{
  "mcpServers": {
    "nano-banana-mcpv2": {
      "command": "docker",
      "args": [
        "run",
        "-i",
        "--rm",
        "-e",
        "GEMINI_API_KEY=your-gemini-api-key-here",
        "nano-banana-mcpv2"
      ]
    }
  }
}

🔧 Available Tools

`generate_image`

Create a new image from a text description using Gemini multimodal native generation.

prompt (required): Description of the image to generate.
model (optional): Custom model name to use for this generation (e.g., gemini-3.1-flash-image).
aspectRatio (optional): Aspect ratio for the image (1:1, 16:9, 9:16, 4:3, 3:4). Defaults to 1:1.

`generate_imagen`

Generate a new high-fidelity image from a text description using Google's dedicated Imagen model (e.g., imagen-4.0-generate-001).

prompt (required): Description of the image to generate.
model (optional): Dedicated Imagen model version (defaults to imagen-4.0-generate-001).
aspectRatio (optional): Aspect ratio for the image (1:1, 16:9, 9:16, 4:3, 3:4). Defaults to 1:1.
numberOfImages (optional): Number of images to generate (1 to 4). Defaults to 1.
negativePrompt (optional): Description of elements to avoid in the generated image.

`edit_image`

Modify a specific existing image file.

imagePath (required): Full local file path of the base image.
prompt (required): Description of modifications.
referenceImages (optional): Array of image file paths for style transfer or guidance.
model (optional): Custom model name to use.

`continue_editing`

Refine the last image generated/edited in the active session.

prompt (required): Description of modification.
referenceImages (optional): Array of reference image file paths.
model (optional): Custom model name to use.

`get_last_image_info`

Check details of the last generated/edited image in the active session (file path, file size, last modified timestamp).

`get_configuration_status`

Verify if the Gemini token is configured and see its origin source.

`configure_gemini_token`

Configure your Gemini API key:

apiKey (required): Your Google AI Studio Gemini API key.

📁 File Storage Directories

Images are saved automatically to:

Windows: %USERPROFILE%\Documents\nano-banana-images\
macOS/Linux: ./generated_imgs/ (or ~/nano-banana-images/ if run from system directories).

🤝 Contributing & Branches

main: Production-ready, stable releases (tagged v*.*.*).
dev: Active features, improvements, and pre-releases (tagged v*.*.*-beta.*).

Make sure to commit changes to the dev branch and open a PR to main for release.

📄 License & Credits

This project is licensed under the MIT License - see the LICENSE file for details.

Acknowledgments

Original Project: Forked from the excellent ConechoAI/Nano-Banana-MCP (originally generated by Claude Code).
Google AI: For the powerful Gemini Multimodal Image APIs.
Anthropic: For the Model Context Protocol (MCP) specification.

MCP Servers