Higgsfield MCP with Claude: a guide for AI ad makers

What is Higgsfield MCP?

Higgsfield MCP is a connector that links your Higgsfield account to Claude or another agent that supports the Model Context Protocol. The agent can then generate images, video, Soul characters and audio, and check your credit balance, without you opening higgsfield.ai (Higgsfield Help Center).

It is not the Higgsfield API. It needs no API key and signs in with OAuth. Every result is saved to your Assets on higgsfield.ai, tagged with the MCP source.

The Higgsfield MCP page lists models including Seedance 2.5, Seedance 2.0, Kling 3, Nano Banana Pro, GPT Image 2 and Soul 2.0, plus prebuilt workflows grouped as production skills, paid ads and motion design.

For ad makers, Claude holds the brief, script and asset list, writes the prompts and sends them to Higgsfield in the same chat, so you stop copying prompts between tabs.

Which tools do creators use with Higgsfield MCP?

Most tutorials pair Claude with Higgsfield, and use Seedance for video and GPT Image or Nano Banana for stills. Based on 17 public tutorials posted between 4 May 2026 and 11 Oct 2026, reviewed by Anglr.

Eight tutorials use the MCP connection itself. The rest have Claude write prompts that the creator pastes into Higgsfield by hand, which is the loop the MCP removes. @mikefutia describes that manual back-and-forth as the step his tool automates.

How do you connect Higgsfield to Claude?

Add a custom connector in Claude with the URL https://mcp.higgsfield.ai/mcp, then sign in to Higgsfield when the authorization window opens. You need an active paid Higgsfield subscription; no API key is needed (Higgsfield Help Center).

Claude on the web or desktop (Pro or Max), per Anthropic's help article:

1. Open Customize, then Connectors. Higgsfield's own article calls this Settings, then Connectors.
2. Click "+ Add", then "Add custom connector".
3. Enter Higgsfield as the name and the server URL above.
4. Click Add or Connect, then sign in and authorize in the Higgsfield window. If you are already signed in on higgsfield.ai, this completes on its own.
5. In a chat, open the + menu, choose Connectors and switch Higgsfield on.

Free Claude plans allow one custom connector. On Team plans only owners can add one; members then click Connect on it. Claude asks before it calls a tool, and Anthropic advises choosing "Allow always" only for tools you trust to run unsupervised.

Claude Code. Higgsfield recommends its CLI and skills for coding agents, because they use fewer tokens than MCP (Higgsfield CLI article):

npm i -g @higgsfield/cli
higgsfield auth login
npx skills add higgsfield-ai/skills

To use the MCP server in Claude Code instead, add it with Anthropic's documented command, then run /mcp inside Claude Code to sign in (Claude Code docs):

claude mcp add --transport http higgsfield https://mcp.higgsfield.ai/mcp

Other agents. ChatGPT uses the Higgsfield plugin instead of the URL, and Cursor adds Higgsfield from its marketplace. Audio generation and the website-building skill are not available in ChatGPT.

Credits. Higgsfield's pricing page, read 12 Oct 2026, says unlimited models and free generations are available only on higgsfield.ai, not through MCP, CLI, Canvas or Supercomputer (Higgsfield pricing). There is no credit cap in MCP. Higgsfield suggests telling the agent to stay under a limit and ask before it generates, and notes this is an instruction, not a hard limit. Plan prices are on Higgsfield pricing.

How do creators use Higgsfield MCP to make ads?

The tutorials agree on one order: research a winning reference, have Claude plan the shots and write the prompts, make the stills before any video, render scenes in batches, then finish by hand. They differ on how much the agent does without a person checking.

1. Start from what already works. Seven authors begin with research or a reference: @kidpakerot, @eptwts and Alber Adilov research trends, @petersapps uses the Meta Ad Library, Tim Scheuer gives Claude other launch videos, and Estefania Freites breaks reference ads down shot by shot. They copy structure, not footage.
2. Give Claude the brief and let it write each shot's prompt. Almost every source shares this step. They differ on input: @MaxbatIe and Maverick Maltin give one short request, while Abhinav Kalyan and Estefania Freites attach the full script and every labelled asset.
3. Approve before credits are spent. Mike Futia approves angles and copy before anything generates, and Alber Adilov approves each batch. Higgsfield's own advice to have the agent ask first points the same way.
4. Make the stills first. Abhinav Kalyan, Estefania Freites and Maneesh Shakya build character, product and location images, and fix problems there, before rendering video. Estefania Freites puts it as "Images are cheap, video is expensive."
5. Render each scene in batches and keep the best take. Abhinav Kalyan runs four or more batches per scene, Mike Futia fires several prompts at once, and Alber Adilov and Maverick Maltin generate ads in bulk from one request.
6. Finish by hand. Maneesh Shakya edits the final cut himself, Estefania Freites assembles in ffmpeg, and Mike Futia keeps all text as live Figma layers: "Higgsfield generates the background. Figma holds the headline, body copy, CTA, and logo as real layers your team can retype."

Two authors go further and let the agent publish. Alber Adilov and Deykhan Ten connect Meta Ads so the agent pushes ads live. If you make ads for brands, have every ad checked against platform and brand rules before it runs.

What should you tell Claude?

Tell Claude the ad, the limits and the prompt structure in one message. Higgsfield's prompt guidance starts with subject, action, setting and camera, then adds lighting, style, dialogue and sound (Higgsfield prompts). The template below is ours; build single-shot prompts with the AI ad prompt builder.

Use the Higgsfield connector for this job.

Ad: [length]-second [format] ad for [product], aspect ratio [ratio].
Script: [paste script, one line per beat]
Assets: I will upload [product photo], [character image] and [location image]
through the Higgsfield upload window. Refer to them as [label 1], [label 2], [label 3].

Rules:
1. First, show me a shot list: shot number, length, what the camera sees, spoken line.
2. Write one prompt per shot in this order: subject, action, setting, camera,
   lighting, style, dialogue (Speaker: "line"), sound.
3. One action per shot. Keep [label 1] identical in every shot.
4. Make stills for shots 1 to [n] with [image model] before any video.
5. Do not spend more than [number] credits. Ask me before every generation.
6. Render video with [video model], [number] takes per shot, and list the takes.

Common problems and fixes

Most failures in the tutorials come from consistency and from long renders; most connection failures have a fix in Higgsfield's help center.

What creators report

These are the authors' own claims, not tested by Anglr.

• Alber Adilov reports that 100 UGC ads took about 15,000 credits and about three hours of compute.
• Maverick Maltin reports getting 50 UGC-style ads from a single request.
• @kidpakerot reports going from hook to finished video in 10 minutes.
• Abhinav Kalyan reports that his 30-second film is the best of about 100 attempts.
• Maneesh Shakya reports a film of almost seven minutes with 17 to 18 hours of human time, and says "The most useful result was not “perfect AI cinema.” It was iteration control."

Volume is the easy part with an agent. The ads that earn money still need a tested script and a clean finish. If your ads already perform, Anglr pays AI creators on performance to remake winning scripts in every format: see what we look for. More tool and style guides are on AI ad guides.