Banana Claude 3.0.0: the full creative system

Banana Claude 3 is no longer a prompt-expansion helper. It is a complete,
plan-first Claude Code plugin for generating, editing, comparing, and reviewing
images with Google's current Gemini image models.

One request can become a frozen visual brief, a visible cost and permission
plan, a bounded set of Gemini calls, private artifacts, and a pixel-level review.
Simple requests remain simple. Complex work can compare up to three prompts
across three model routes without surrendering the approval boundary.

Install

Requirements: Claude Code 2.1.199 or newer, Python 3.11 or newer with python3
on PATH, and a Gemini API key for a billing-enabled Google AI project.

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

Then try:

/banana-claude:banana generate an editorial 16:9 campaign hero with clean copy space

Highlights

• Current routes for Nano Banana 2 Lite, Nano Banana 2, and Nano Banana Pro,
  plus one clearly marked deprecated compatibility route.
• A bundled, zero-dependency Python MCP server with no unpinned third-party
  provider package.
• Brief-led generation, preservation-aware editing, stored continuation on
  Flash and Pro, Search grounding, Flash video-to-image, and bounded portfolios.
• Canonical visual briefs and 30-minute, single-use approvals bound to the exact
  prompt, inputs, route, estimate, output, and privacy settings.
• One provider attempt per approval, atomic private artifact publication, and
  concurrency-safe cost reconciliation.
• Deterministic exact-copy SVG composition for approved text, fonts, logos, and
  trusted raster artwork.
• Optional read-only visual architect and critic agents, with the human approval
  boundary kept outside experimental agent orchestration.
• Offline migrations, a redacted legacy-install scanner, a private doctor, and
  a backup-first standalone installer that refuses unowned destinations.
• Deterministic tests for request validation, provider parsing, approvals,
  concurrency, outputs, presets, migrations, installer lifecycle, and repository
  contracts.

Breaking upgrade note

Public 1.4.1 and 2.1.0 installs may contain an unpinned third-party MCP entry, a
raw Google key in Claude settings, an obsolete skill directory, or a key-bearing
shell command. Installing version 3 does not remove that residue.

Before enabling version 3, read the
upgrade guide, run the
redacted legacy scan, review any remediation fingerprint, and revoke or rotate a
key that was stored by an older setup. Never post a key, settings backup, private
image, or full environment dump in an issue.

State schemas also changed. Version 3 provides explicit dry-run and
fingerprint-confirmed migrations for exact legacy cost ledgers and presets. It
does not rewrite them during install or normal loading.

Verification

The local release candidate passed:

• 387 deterministic offline tests on Python 3.11 and Python 3.14.
• Ruff lint and formatting checks across 27 Python files.
• Strict Mypy across the same 27 Python files.
• Python compilation, Bash syntax, ShellCheck, strict Claude plugin validation,
  repository contract checks, and git diff --check.
• A disposable-home install, doctor, and recoverable uninstall lifecycle.
• A candidate-file secret scan with zero findings.
• A pinned development-dependency audit with no known vulnerabilities.

Private ignored receipts from the August 28 paid acceptance run were rechecked
against their sidecar hashes. Pro, Lite, and deterministic typesetting passed
visual review. The original Flash image and its preservation edit remain honest
negative QA examples and are not presented as passing showcase work. The
release is published only after public GitHub CI passes on its exact commit.

Thanks

Thanks to @qant-au for independently reporting the
legacy cost-ledger concurrency risk in pull request 9. Version 3 resolves the
same underlying race through its broader locked, atomic, versioned ledger and
concurrent-writer test coverage.

Thanks to every user who starred the project, opened a report, proposed an idea,
or used Banana Claude in real creative work. This release exists because the
small original skill deserved a serious second life.

Known limits

• Gemini API image inference has no free tier on these current routes.
• Generated consistency, exact spelling, geometry, policy acceptance, rights,
  and production fitness still require review.
• Banana's estimates are not invoices or hard provider spending caps.
• JPEG is the only selectable provider output MIME in this release. PNG remains
  supported for input and trusted local rendering.
• Live provider calls, pixel acceptance, destination-platform checks, and rights
  review remain separate human gates and never run in public CI.

Read the
full changelog
for every breaking, added, changed, fixed, and security item.