nano-banana-mcp

A minimal MCP server that wraps Google Gemini image generation ("Nano Banana") for use inside Claude Code or any other MCP-compatible host.

Built as part of an exploration of agentic workflows for K-12 teaching materials. The whole server is about 75 lines of Python — readable in one sitting.

What it does

Exposes a single tool:

generate_image(prompt: str, output_path: str, model: str | None = None) -> str

• Writes a PNG to output_path

• Auto-suffixes on collision (foo.png → foo-1.png, foo-2.png, …) — never silently overwrites

• Returns the absolute path it actually wrote to

Install on a new machine

You need:

• Homebrew (macOS)

• A Gemini API key from Google AI Studio

• Claude Code (or another MCP host)

brew install pipx
pipx ensurepath
pipx install git+https://github.com/DenialGelon/nano-banana-mcp

claude mcp add nano-banana --scope user \
  --env GEMINI_API_KEY=<your-key> -- nano-banana-mcp

Verify with claude mcp list — you should see nano-banana: ✓ Connected.

Use

Start a fresh Claude Code session and ask:

Generate an image of a friendly cartoon banana mascot and save it to ~/Desktop/banana.png

Claude will pick up the tool and call it. The return value tells you the final path (with a note if a collision was auto-resolved).

Models

The model argument is optional. Known values (verified 2026-05-18 at the Gemini image-generation docs):

Update / rotate the key

Update the code:

pipx install --force git+https://github.com/DenialGelon/nano-banana-mcp

Rotate the API key:

claude mcp remove nano-banana --scope user
claude mcp add nano-banana --scope user \
  --env GEMINI_API_KEY=<new-key> -- nano-banana-mcp

License

MIT — see LICENSE.