Claude skill install
The nano-banana Claude skill teaches Claude how to drive the MCP server
fluently — when to architect first, when to one-shot, how concept_id
threads brand styling through, how to surface progress to a human.
Without the skill, Claude still calls the tools, but defaults can be
suboptimal (e.g., one-shot a raw prompt with no concept proposal, or skip
showing progress updates).
Install
Vendoring locally (Claude Code / Desktop)
git clone git@git.cpl.cloud:Intern/ai/genkit/nano-banana.git /tmp/nano-banana
cp -r /tmp/nano-banana/skills/nano-banana ~/.claude/skills/
Restart Claude Code. The skill is auto-discovered; verify with /skills.
Antigravity CLI / agentskills.io
Coming soon: published skill package on agentskills.io.
Register the MCP server
Independently of the skill, register the remote MCP at
https://mcp.nano.cpl.ai in your client:
Claude Code:
claude mcp add --transport http nano-banana https://mcp.nano.cpl.ai/mcp
Then run /mcp auth nano-banana to complete the OAuth flow. Sign in with
your @cloudpilots.com Google account.
Claude Desktop / claude.ai: Settings → Connectors → "Add custom
connector" → URL https://mcp.nano.cpl.ai/mcp.
What the skill does
The skill is a folder of Markdown files:
skills/nano-banana/
├── SKILL.md # Trigger phrases, MCP registration check, workflow
├── references/
│ ├── tools.md # Tool contracts + arg shapes (mirror of this site)
│ ├── workflows.md # Decision tree: when to architect vs one-shot
│ └── styling.md # How concept_id carries the brand-styled prompt
└── VERSION
When Claude sees a request like "generate a marketing image of X" or "show me some options for a hero illustration", the skill activates and walks Claude through:
- Setup check — verify the four MCP tools are present, surface a clear error if not.
- Workflow decision — propose concepts first vs one-shot, based on how concrete the user's prompt is.
- Generate — call the right tool with the right args; pass the
concept_idwhen one was chosen so brand styling isn't lost. - Surface progress — for synchronous calls, let Claude relay the
per-stage progress notifications; for async, suggest a follow-up
get_imagepoll.
See Brand styling for the why behind the workflow split.