Generate an image in Claude with a remote MCP connector

Connect Claude Imagine to Claude, check the price of one image, then generate it with a credit cap. The example here is a product shot for a fictional skincare brand called LUMA. In our September 5 Claude Desktop test, we asked for a vertical product image using Nano Banana 2. The request cost 3 credits, and the image appeared in an embedded card in the conversation. Disclosure: I build Claude Imagine, the independent service used here. It isn't affiliated with Anthropic. AI-generated example from our September 5 test. LUMA is a fictional product. This is the output image, not a screenshot of Claude. Connect the service Claude handles the conversation and the tool request. Claude Imagine sends the image request to an image-generation model. You'll need accounts for both services; their usage balances are separate. In Claude, find Customize → Connectors, choose Add custom connector, and enter: https://claudeimagine.com/api/mcp Name it Claude Imagine. Add the connector, then use Connect if authentication hasn't started. Sign in to Claude Imagine and approve the connection. Enable it for the conversation if necessary. Menu labels can change. Anthropic maintains the current custom-connector instructions, including organization setup. Our September 5 generation test used an already connected Claude Desktop account; it didn't repeat account registration. Ask for the cost first Start with a request that doesn't create anything: Use Claude Imagine to list the available image models and check my balance. Quote one nano-banana-2 text-to-image generation at 9:16. Don't generate an image yet. The relevant tools are list_models, get_credits, and quote_generation. The quote should tell you the cost and whether your account can use the requested model. Our recorded image cost 3 credits. Treat that as the cost of that test, not a permanent price promise. Check the quote returned for your account before continuing. Generate one image Here is a starter prompt for the same type of task. It's a reusable brief, not a transcript of the original test: Generate exactly one image using Claude Imagine and nano-banana-2 at 9:16. Set max_credits to 3. If the request costs more, stop. Create a product image for a fictional skincare brand named LUMA. Place one bottle on a stone plinth, with a lime nearby and soft studio light. Use a simple background and leave space above the bottle for a future headline. Don't add a headline to the image. Show the result using the available image card and include the original file link. Don't generate alternatives or video. Approve the tool call if Claude asks. The max_credits value caps this individual request; it doesn't create a shared budget for future edits. If the quote exceeds the cap, choose a different model or deliberately change your budget. Check the result you actually received In our Desktop test, the result appeared in an embedded card with an original-image link. That observation doesn't establish the same rendering behavior in Claude Code, mobile, or every web-client version. On September 8, we ran a separate English-language test in claude.ai: one new LUMA image with Nano Banana 2 at 16:9. Claude quoted 3 credits before generation and reported the same cost afterward. The finished image appeared in the conversation. Actual claude.ai screenshot, September 8. This AI-generated 16:9 image is a separate result from the September 5 vertical image above. The prompt shown in the card differs from the starter template. The screenshot is cropped to leave out private account details. Check the tool result if you don't immediately see the preview. Keep the original link as well. If the file exists but the preview is missing, don't immediately ask Claude to generate again: retrieving a result and creating a second image are different operations. Inspect the image before using it. Product labels, bottle shape, and spacing need a visual check. A completed tool call confirms that a file was produced; it doesn't confirm that the image meets your brief. Once the first image works, you can try a separate edit. In our test, changing the background used edit_image with the previous result's address, so we didn't download and re-upload the image. That edit was another paid operation. The Claude Imagine website has the connector setup and account access. Save the first result before expanding the workflow.