Guide: Generating Product Images Inside Claude with Nano Banana

Guide: Generating Product Images Inside Claude with Nano Banana

Connecting Claude to Google's Gemini image models (Nano Banana 2 and Nano Banana Pro) means you can generate and edit product imagery inside a chat, rather than bouncing between tools and re-explaining your brand every time. The setup takes around fifteen minutes.

But the connector that makes this possible is a community MCP server — built by an individual, installed by pasting code into a config file on your own machine. It is not an official Anthropic or Google connector, and no such connector exists. There are several competing community versions with different maintainers.

That means the install steps are the easy part. The part that actually protects your business is knowing how to evaluate a server before it touches your machine. This guide covers both, and it covers the vetting first — because that's the skill that survives after this specific tool is obsolete.

What You'll Need

Software

• Claude Desktop app — not claude.ai in the browser. This is a local server that runs on your machine, so the browser cannot reach it.

• uv (the Python runner). This server runs on uv/uvx, not Node.js. Most MCP guides online assume Node — ignore them here.

Access

• A Gemini API key from Google AI Studio

Cost Expectations

You are charged per image generated, by Google. This is completely separate from your Claude subscription. Ask for eight variations, you pay for eight images. Check current per-image pricing on Google's Gemini API pricing page and set a spend cap in Google Cloud before you use this at any volume.

The Server

zhongweili/nanobanana-mcp-server — https://github.com/zhongweili/nanobanana-mcp-server

Step 1: Work Out Which Tier You're Dealing With

Not all MCPs carry the same risk, and the difference is structural rather than a matter of taste.

Nano Banana connectors are tier three. Every option is community-built. That's exactly why you need a method rather than a recommendation.

Step 2: Run the Eight Questions

Put every community MCP through these before it goes near your machine.

Who built it?

1. Is there a named human or company behind it? An anonymous account with no history is a hard no.

2. Is it actually maintained? Check the commit history and releases. Something last touched fourteen months ago is abandoned — and abandoned code doesn't get security patches.

3. Is anyone else using it? Stars, forks, download volume, and open issues that actually get answered. Twelve downloads and zero issues means you are the QA department.

What does it want?

4. What is it asking for access to? An image server should want an image API key. If it also wants your file system, your email, or your browser — stop. Permission scope creep is the loudest alarm bell there is.

5. Can you read the code? Open source with a permissive licence (MIT, Apache) means the code is inspectable. You may never read it yourself, but thousands of people can, and that's the point.

6. Does it fetch anything from an external link at runtime? This is the one that catches people. If the server pulls code or instructions from a remote URL, whatever sits behind that URL can be swapped after you install it. It passes every security scanner on day one and turns malicious on day ninety.

What happens if it goes wrong?

7. What's the blast radius? If this server were compromised tomorrow, what could it reach? An image generator with one API key and a folder to write to is a contained problem. An MCP with write access to Seller Central or your customer database is not.

8. Can you contain it? Use a dedicated API key with a spend cap. Don't reuse a key you've already given to five other tools. Assume you may need to revoke it in a hurry.

Step 3: Score the Server — and Decide for Yourself

Here's how zhongweili/nanobanana-mcp-server scores against those checks:

And yet: nobody can guarantee its safety, and that includes anyone recommending it to you. Passing the checks above on a given day is not a security audit. Repos change hands. Maintainers lose interest. Dependencies get compromised upstream. A server that's clean today can be a problem in six months, with no notification to you.

So run the checks yourself before you install, and run them again periodically. Open the repo. Read the recent commits. Scan the open issues. If something has changed, don't install it — and don't install it purely because someone else's guide told you to.

If you're not comfortable, don't install it

There is no obligation here, and two perfectly good alternatives exist:

• Use Claude's built-in design and image capability. For on-brand text carousels and graphics it's often better and faster anyway.

• Have Claude write the image prompt, then run it yourself in Google AI Studio or Gemini. You lose the automation, you keep the thinking, and you install nothing.

"I decided the risk wasn't worth it" is a legitimate, professional answer — and a better one than installing something you don't understand.

Step 4: Install the Prerequisites

A) Check for uv

Open your terminal (Mac: Terminal. Windows: PowerShell) and run:

uv --version

If that errors, install it.

Windows (PowerShell):

powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"

Mac:

curl -LsSf https://astral.sh/uv/install.sh | sh

Close and reopen the terminal, then run uv --version again to confirm.

B) Get a Gemini API key

1. Go to aistudio.google.com → sign in → Get API key → Create API key.

2. Copy it into a password manager. It's shown properly only once.

3. Set a spend cap in Google Cloud before you generate at volume.

Step 5: Add the Server to Claude Desktop

A) Open the config file

Windows: Press Win + R → paste %APPDATA%\Claude → Enter → open claude_desktop_config.json (right-click → Open with → Notepad or VS Code).

Mac: Finder → Go → Go to Folder → paste ~/Library/Application Support/Claude → open claude_desktop_config.json.

If the file doesn't exist, create it with exactly that name.

B) Paste the config

If the file is empty or new, paste this in full:

json

{"mcpServers":{"nanobanana":{"command":"uvx","args":["nanobanana-mcp-server@latest"],"env":{"GEMINI_API_KEY":"PASTE_YOUR_KEY_HERE","IMAGE_OUTPUT_DIR":"PASTE_YOUR_FOLDER_PATH_HERE"}}}}

Two fields to fill in:

• GEMINI_API_KEY — your key from Step 4B.

• IMAGE_OUTPUT_DIR — where generated images get saved. This is optional; delete the line and images land in ~/nanobanana-images. Set it deliberately so you always know where your assets are.

  • Windows example: "C:\\Users\\YourName\\Documents\\NanoBanana" — note the double backslashes. JSON requires them.

  • Mac example: "/Users/YourName/Documents/NanoBanana"

If your config already contains other MCP servers, add nanobanana as a new entry inside the existing "mcpServers": { ... } block, with a comma after the previous entry. Do not create a second mcpServers block.

C) Don't hand-wrestle the JSON

JSON breaks on a single missing comma. If any of that felt fiddly, let Claude do it:

Add a nanobanana MCP server entry to this config using uvx and this API key. Return the complete, valid JSON.

Paste your current config in alongside it, then paste the result back into the file. Faster and less error-prone than counting brackets.

D) Save and fully restart Claude

Save the file, then quit Claude Desktop entirely — not just close the window. Mac: Cmd + Q. Windows: quit from the system tray. Then reopen.

Step 6: Verify the Connection

1. In Claude Desktop, open Settings → Connectors (or the tools icon in the chat box).

2. You should see nanobanana listed with three tools: generate_image, edit_image, upload_file.

If it isn't there, the cause is almost always invalid JSON or a Claude instance that wasn't fully quit.

Step 7: Turn Off Competing Image Tools (Do This Before Testing)

If Higgsfield, Canva, or any other image MCP is switched on, Claude has to guess which one you meant — and it will sometimes guess wrong.

Toggle every other image-generating MCP off. They stay installed; they just leave the room. Switch them back on when you specifically want them.

Fewer tools enabled means fewer misfires, fewer tokens, and more predictable behaviour.

Step 8: Run Your First Test

Open a fresh chat and start deliberately dumb:

Generate an image of a flying kitten.

Watch the tool call fire. You should see nanobanana in the tool trace and get an image back, saved to your output folder.

Then try something real:

Generate a 3:4 lifestyle image of a ceramic coffee mug on a marble kitchen counter, morning light, minimal props.

The rest of the SOP — including the three model tiers and how your word choice silently routes you to the more expensive one, the full aspect ratio list, the complete troubleshooting table (including the Windows uvx PATH fix), how to wrap this in a Skill with a human approval gate, and the quarterly hygiene checklist — is here. [A free gift from me :)]

• TikTok lets brands build and publish episodic microdramas in-app, then amplify them through Growth Max ads. The format hit $1.3B in US viewer payments in 2025, and TikTok claims campaigns drive 52% more incremental reach.

• TikTok launches Agentic Hub, a marketplace of first-party and third-party AI Skills built on its Business MCP. AI agents handle campaign creation, creative generation, and performance diagnostics. Partners include HubSpot and Wix.

• Google tests a dedicated Gemini inbox for Workspace users, with filters to triage, track, and review tasks. A "Needs review" filter points at a proactive agent that pulls email from connected sources into a structured to-do list.

• Digitas North America CEO Amy Lanzi calls AI the new programmatic hype at Cannes, warning that "free AI" pitch promises hurt the industry. Parent Publicis ran an ad mocking false AI claims. She says the traditional CMO role is finished.

• Google's new paper details S-CTS, a system that terminates whole clusters of coordinated accounts mass-producing templated AI slop, not single pages. Enforcement shifts from grading the artifact to the behavior behind it, closing arbitrage.

• Anthropic rolls out Claude Cowork to mobile and web, starting with Max users. Sessions and files sync across devices, scheduled tasks run with no device online, and over 90% of Cowork use is knowledge work, not coding.

• OpenAI launches ChatGPT Work, a Codex-powered agent that builds sheets, slides, docs, and apps across connected tools. It runs multi-step projects for hours and continues via Scheduled Tasks. Nearly 100% of internal OpenAI teams use it.

• OpenAI releases the GPT-5.6 family: flagship Sol, balanced Terra, and low-cost Luna. Sol tops Agents' Last Exam at 53.6, beating Claude Fable 5 by 13.1 points, and leads coding benchmarks with fewer tokens at lower cost.

📊 Dataku - Extract valuable insights from documents and texts.

🥇 Outrank - Find untapped keywords and automatically publishes daily SEO-optimized blog posts to grow your website's traffic

🤖 Forefront - Your AI assistant for work

🔎 Agora - The search engine for e-commerce products.

🎙Outcast - AI content creation for podcasts

😍It was great, all I needed

🙂Good, not great

🙄It sucked