Nano Banana Claude Skill: The Real Repos and How They Work

A "Claude banana skill" is a community-built Claude Code skill that lets Claude generate and edit images with Google's Nano Banana models. Anthropic doesn't ship one, and neither does Google. Every version you'll find is a GitHub repo written by an independent developer.

They all work the same way under the hood:

• A SKILL.md file tells Claude when to reach for the skill and what to run.
• A small script sends your prompt to the Gemini API and saves the image that comes back.
• Your own Gemini API key pays for each image. Google's pricing page lists no free tier for the Nano Banana models.

Below: the three repos worth knowing, what each one's README actually documents, and the one setup detail that trips people up most, which is model IDs that have already been shut down.

How a Nano Banana skill works in Claude

A skill is a folder of instructions Claude can load on demand. Anthropic's skills documentation puts it simply: "Create a SKILL.md file with instructions, and Claude adds it to its toolkit." If skills are new to you, our Claude Skills guide covers the basics.

A Nano Banana skill has three moving parts:

Claude only reads the full skill when it needs it. The docs note that "full skill content only loads when invoked," so a long image skill costs almost nothing in sessions where you never ask for a picture.

Personal skills go in ~/.claude/skills/<skill-name>/SKILL.md and load in every project. Skills that come inside a plugin are namespaced, so you call them as /plugin-name:skill-name.

The script is the part that talks to Google. Every Nano Banana skill needs a Gemini API key, which you create in Google AI Studio. Google's API key guide says to "Set the environment variable GEMINI_API_KEY or GOOGLE_API_KEY," and the official client libraries pick it up automatically.

Banana Claude: the full plugin

The repo most people mean by "Claude banana" is AgriciDaniel/banana-claude (the old claude-banana URL redirects there). It's MIT-licensed and is the most elaborate of the three: a Claude Code plugin that bundles a skill, two helper agents, and a local MCP server.

Its README sums up the job as "Create, edit, compare, and review Gemini images from Claude Code." What stands out is how cautious it is with spending:

• Plans offline first. You see the exact prompt, model, output path, and a cost estimate before any paid call.
• One approval, one attempt. A changed plan, a retry, or a regeneration needs a fresh approval. Its SKILL.md tells Claude never to silently auto-retry.
• Installs disabled. Because image generation costs money, you have to switch the plugin on yourself.
• Keeps the key out of chat. The plugin asks for your Google key as sensitive config and hands it only to its bundled MCP process.

It routes between Nano Banana 2.1 (the default), Nano Banana 2 Lite, and Nano Banana Pro. The README also keeps a Nano Banana 2 route for 512px drafts and calls that model deprecated, which matches Google: the Gemini API changelog marked gemini-3.1-flash-image deprecated on October 6, 2026, with no shutdown date announced, and points new projects to Nano Banana 2.1. It also does edits, side-by-side prompt comparisons, and local typesetting when you need exact text on an image.

Requirements, per the README: Claude Code 2.1.199 or newer, Python 3.11 or newer, and a Gemini API key on a billing-enabled Google AI project. Generation runs on Linux and macOS. Native Windows isn't supported yet, so the README points Windows users to WSL.

The install commands, copied from the README:

/plugin marketplace add AgriciDaniel/banana-claude
/plugin install banana-claude@banana-claude-marketplace
/plugin enable banana-claude@banana-claude-marketplace
/reload-plugins

Then ask in plain language, for example /banana-claude:banana generate an urban 16:9 GitHub hero with clean left-side copy space.

The repo also offers a standalone skill install through its install.sh script. That version gives you /banana and the scripts, without the MCP server or the managed key setup.

nanobanana-skill: the lightweight option

If you want the classic "SKILL.md plus one Python script" shape, look at nanobanana-skill inside feiskyer/claude-code-settings, an MIT-licensed collection of Claude Code skills and agents. The folder holds just three files: SKILL.md, nanobanana.py, and requirements.txt.

What the skill documents:

• Text-to-image and editing. Pass --input with one or more images to edit them.
• Sizes and resolution. Ten aspect ratios via --size (9:16 portrait by default) and --resolution of 1K, 2K, or 4K.
• Search grounding on by default. --no-search turns it off. --no-think disables thinking mode.
• Key setup. GEMINI_API_KEY, either exported or saved in ~/.nanobanana.env.
• Dependencies. google-genai, Pillow, python-dotenv, and httpx, installed from requirements.txt.

The repo's README recommends installing through the Claude Code plugin marketplace:

/plugin marketplace add feiskyer/claude-code-settings
/plugin install claude-code-settings

That main plugin installs the whole collection, not just the image skill. The README also lists npx skills for picking individual skills, with a warning that versions installed that way can lag behind the repo.

One catch to fix before your first image. The skill's default model is gemini-3.1-flash-image-preview, and its other listed option is gemini-3-pro-image-preview. Google's deprecations page marks both preview IDs as shut down, with a shutdown date of June 25, 2026. The script accepts any model string, so pass a current ID with --model (see the table below) and confirm the call succeeds.

cc-nano-banana: the Gemini CLI route

kkoppenhaver/cc-nano-banana takes a different path. Its SKILL.md doesn't call the Gemini API from its own script. Instead, it has Claude drive the Gemini CLI's nanobanana extension.

That means more setup. The README walks through installing the Gemini CLI, setting NANOBANANA_GEMINI_API_KEY, installing the nanobanana extension, and then copying SKILL.md into ~/.claude/skills/nano-banana/. In return you get the extension's commands, including /generate, /edit, /restore, /icon, /diagram, /pattern, and /story. Images land in ./nanobanana-output/.

Check the model before you rely on it. The README names gemini-2.5-flash-image as the default, which Google now labels its legacy Nano Banana model. The README's suggested upgrade, gemini-3-pro-image-preview, is one of the preview IDs on Google's shutdown list.

Quick comparison of the three repos

Which Nano Banana model IDs work now

Model IDs are the part of these skills that goes stale fastest. Here's how Google's image generation docs list the family as of October 28, 2026:

One inconsistency to know about: Google's pricing page still carries a warning that gemini-2.5-flash-image shuts down on October 2, 2026, while the deprecations page gives March 15, 2027. Either way, Google recommends moving to Nano Banana 2 Lite, so don't build anything new on it.

Google's docs also say "All generated images include a SynthID watermark," so that applies whichever skill you use. For how the Pro and Flash-tier models differ in practice, see our Nano Banana 2 vs Pro comparison.

On cost: Google's Gemini API pricing page shows "Not available" in the free tier column for every Nano Banana model. Each image a skill generates is billed to your project, and the rates are on that page. For what you can and can't get without paying, see is Nano Banana free?.

Build your own Nano Banana skill

None of the repos fit? A minimal skill is two files. This is a sketch, not a tested package.

1. ~/.claude/skills/nano-banana/SKILL.md

---
name: nano-banana
description: Generate an image with Google's Nano Banana model. Use when the user asks to create or generate an image.
---

Run `python3 ${CLAUDE_SKILL_DIR}/generate.py "<prompt>"` with the user's
description as the prompt, then tell the user where the image was saved.

${CLAUDE_SKILL_DIR} is a variable Claude Code fills in with the skill's own folder, so the script path works from any project.

2. ~/.claude/skills/nano-banana/generate.py, based on the Python example in Google's image generation docs:

import base64
import sys
from google import genai

client = genai.Client()  # reads GEMINI_API_KEY from the environment

interaction = client.interactions.create(
    model="gemini-nano-banana-2.1",
    input=sys.argv[1],
)

with open("generated_image.png", "wb") as f:
    f.write(base64.b64decode(interaction.output_image.data))

Install the SDK with pip install google-genai, export your key, and ask Claude for an image. Once that works, you can add editing, aspect ratios, or an approval step the way Banana Claude does.

FAQ

Is there an official Nano Banana skill for Claude?

No. Anthropic and Google don't publish one. Banana Claude, nanobanana-skill, and cc-nano-banana are all community projects. Read the scripts before you install any of them, since they run on your machine with your API key.

Do I need a Gemini API key for a Claude banana skill?

Yes. Each skill sends your prompt to Google's Gemini API, so it needs a key from Google AI Studio. Banana Claude's guide specifically asks for a key on a billing-enabled project.

Does the Nano Banana skill for Claude cost money?

The three skills are MIT-licensed and cost nothing to install, but every image they make is a paid Gemini API call. Google's pricing page lists no free tier for these models. Banana Claude shows an estimate and asks for approval before each call; the other two don't describe a spend check.

Can Claude make images without a skill?

Claude's models take images as input but don't output them. A skill or an MCP server is how Claude Code hands image work to a model like Nano Banana. If you're weighing Google's model against OpenAI's, our GPT Image 2 vs Nano Banana comparison breaks down the trade-offs.

Why does my nano banana skill return a model-not-found error?

The skill is probably still pointing at a preview ID Google has shut down, such as gemini-3.1-flash-image-preview or gemini-3-pro-image-preview. Switch it to a current ID from the table above.

Key takeaways

• A Claude banana skill is a SKILL.md plus a script that calls Google's Gemini API. All of them are community-built.
• Banana Claude is the most complete option, with offline planning and an approval before every paid call.
• nanobanana-skill is the simplest to read and modify, but its default model ID has been shut down.
• cc-nano-banana goes through the Gemini CLI and defaults to Google's legacy model.
• Whichever you pick, you need a Gemini API key, and every image is billed to your Google project.

Getting an image skill working is mostly setup: the right repo, the right key, the right model ID. If that setup step is where you keep stalling, Taku mirrors a working AI setup into a desktop workspace and runs it, so you can use a power user's configuration without rebuilding their environment first. Taku is in Beta, and the Mac app is available now.