Claude + Higgsfield: AI Marketing Assets in Minutes
In this walkthrough I turn Claude into an all-in-one AI marketing hub by connecting it to Higgsfield through their brand-new MCP server. Higgsfield gives you access to the state-of-the-art generation models: GPT Image 2 from OpenAI, Nano Banana 2 from Google, plus Seedance 2.0 and Kling 3.0 for video. Claude handles the prompting, the formats, and the model selection, so you go from idea to usable campaign assets in minutes.
What you need
- A Claude account where you can add custom connectors
- A Higgsfield account for the MCP server URL and the generations
- Optional: a speech-to-text tool like WisprFlow if you prefer dictating prompts (watch for transcription quirks, covered below)
1. Connect the Higgsfield MCP to Claude (0:51)
On the MCP tab of the Higgsfield homepage, copy the server URL. Then in Claude: click your profile, go to Settings, head into Connectors, scroll all the way to the bottom, and click Add Custom Connector. Put the name, Higgsfield, and the server URL, click Add, connect your account once, and it is done.
2. Check the MCP connection (1:32)
Before generating anything, verify the wiring:
Hey, Claude, are you connected to the Higgsfield MCP server?
Claude scans the tools it has access to and confirms whether everything is working. Remember that MCP connections can simply fail when the server is down, so this quick check saves you a confusing session later.
3. Generate two product images with two models (2:09)
Now the fun part. I asked for the same product from two different models at the same time:
I've been working a lot on my new brand of protein powder called Berserk Powder. What I want you to do is I want you to create two new images, one with GPT Image 2 and the other one with Nano Banana 2, showcasing the product. The branding is dark with red accents.
Claude first verifies that the models are available on the Higgsfield platform, then creates the two images simultaneously.
4. Why running this through Claude beats doing it yourself (2:50)
The added value of using Higgsfield through Claude is that Claude prompts the models for you. It handles multiple generations at once across several models and manages the image editing loop. It even chose the one-to-one square format on its own because we said the images were for social media, which works great for LinkedIn or Instagram. Claude is also fantastic for brainstorming brands and visuals, so you can have a long discussion about angles before generating. The Nano Banana 2 result looked very good for something generated in fifteen seconds.
5. Turn the best image into UGC videos (4:45)
I told Claude the second image was perfect and asked for videos from the two models people love right now:
Perfect. Now what I want you to do is I want you to take the Nano Banana 2 image, and I want you to create two different videos of about eight to ten seconds using Kling 3.0 and Seedance 2.0. It must be an UGC format with a guy showing the protein powder and how he dry scoops it.
Claude drafted a vertical UGC TikTok-style video: camera starts on the product, then a muscular bearded man in a tank top steps into the frame, 9:16, ten seconds, using the Nano Banana image as the visual reference. Video generation normally takes about thirty seconds to one minute, more with the bigger pro versions of Kling 3.0 and Seedance 2.0.
6. Fix speech-to-text model names (5:36)
I dictated that prompt with WisprFlow, and it wrote "C9 2.0" instead of Seedance 2.0. Claude is normally smart enough to understand what you meant, and when it is not sure it will come back and ask what you mean by that model. Always verify what you are actually sending, but do not stress over small speech-to-text slips; the models today usually match the intended name.
7. Review the results and find your assets (8:33)
About two minutes later I had both videos: a UGC-style clip of a guy dry scooping the powder, and a second video with cuts that look edited by a human, which is just mind-blowing. Seedance 2.0 takes the crown today. And if you ever worry about losing a conversation with Claude, everything lives in the Assets folder on Higgsfield, where you can also check whether a slow generation has finished. Nothing gets lost.
8. Scale it up (9:23)
You can create up to four videos in parallel and keep only what you like, and you can generate 4K videos when you need them. The next step is turning the volume up: hundreds of image or video variations for the same campaign, then measuring which ones perform better.
Pitfalls and tips
- MCP servers can be down. Ask Claude to confirm the connection before you start a session
- A generated image sometimes fails to display in Claude even though it was created. Check the Higgsfield Assets folder before regenerating
- Dictating prompts? Verify model names before sending; speech-to-text mangles names like Seedance
- This feature is brand-new, so take it with a grain of salt and expect the occasional display bug
Where to go next
If you want to go deeper on what the newest image models can do for a business, watch Get Your Money's Worth from ChatGPT Images 2.0. To keep all these assets on-brand instead of one-off pretty pictures, read Design Systems for AI: The 2026 Guide.
