nanobanana-skill
A Claude Agent Skill for image generation, editing, and restoration via Google's Nano Banana (Gemini image models).
When you ask Claude things like "generate a watercolor of a fox", "make a favicon for my app", "edit this photo to add sunglasses", or "draw an architecture diagram for a microservices system", this skill activates and runs the appropriate Gemini image model. Default: gemini-3.1-flash-image (Nano Banana 2).
The repo is also a self-contained Claude Code plugin marketplace — install with two commands (no venv, no pip install).
Install (recommended — as a plugin)
In Claude Code:
/plugin marketplace add batou9150/nanobanana-skill
/plugin install nanobanana@batou9150-skills
Then set the API key (see API key — where to put it) and you're done.
The CLI is self-contained: it uses a PEP 723 inline-metadata shebang to auto-install google-genai on first run via uv. No manual venv or pip install step.
Requirements
uv on PATH (one-line install: curl -LsSf https://astral.sh/uv/install.sh | sh)- A Gemini API key
Install (alternative — bare skill, no plugin system)
git clone https://github.com/batou9150/nanobanana-skill /tmp/nanobanana-skill
mv /tmp/nanobanana-skill/skills/nanobanana ~/.claude/skills/nanobanana
That's it. The first time the script runs, uv resolves and caches google-genai automatically.
API key — where to put it
Get a key from https://aistudio.google.com/apikey, then choose one of:
Option A — ~/.zshenv (recommended for zsh users)
echo 'export NANOBANANA_API_KEY="<your-key>"' >> ~/.zshenv
~/.zshenv loads for every zsh invocation, including the non-interactive shells Claude Code's Bash tool uses. Putting it in ~/.zshrc only works for interactive terminals — Claude Code won't see it.
(Bash users: use ~/.bash_profile or whichever rc your shell sources for non-interactive runs.)
Option B — Claude Code settings only
Add to ~/.claude/settings.json:
{
"env": { "NANOBANANA_API_KEY": "<your-key>" }
}
Stored in plain text on disk; only visible to Claude Code processes (not to standalone CLI use).
Fallback env-var names
If you already have a Gemini key set under another name, the skill picks the first that's defined, in this order:
NANOBANANA_API_KEY (preferred)NANOBANANA_GEMINI_API_KEYNANOBANANA_GOOGLE_API_KEYGEMINI_API_KEYGOOGLE_API_KEY
Subcommands
| Command | Purpose |
|---|
generate | Text-to-image with optional batch styles / variations |
edit | Modify an existing image |
restore | Repair / enhance an image |
icon | App icon, favicon, UI element |
pattern | Seamless pattern / texture |
story | Sequential frames (story / process / tutorial / timeline) |
diagram | Flowchart, architecture, schema, wireframe |
Direct CLI use (without Claude)
The script is fully standalone:
~/.claude/skills/nanobanana/scripts/nanobanana.py generate \
"sunset over mountains" --count=3 --styles=watercolor,oil-painting
Output goes to ./nanobanana-output/ in your cwd.
Run any subcommand with --help to see flags:
~/.claude/skills/nanobanana/scripts/nanobanana.py generate --help
Model override
export NANOBANANA_MODEL=gemini-3-pro-image # Nano Banana Pro
export NANOBANANA_MODEL=gemini-2.5-flash-image # Nano Banana v1
Credit
Adapted from the official gemini-cli-extensions/nanobanana Gemini CLI extension. This is a Claude-Skill port — no MCP server, no Node runtime, just a self-contained Python script invoked from SKILL.md.
License
MIT — see LICENSE.
