Higgsfield MCP: Generating Video From an AI Assistant

Higgsfield MCP is Higgsfield’s own hosted Model Context Protocol server. It lets an AI assistant such as Claude, ChatGPT or Claude Code generate images and short videos through your Higgsfield account, using the same models and the same credits as the Higgsfield website. You connect it with a URL and sign in with your Higgsfield account; there is no API key to copy. As of September 30, 2026, the server address on Higgsfield’s MCP page is https://mcp.higgsfield.ai/mcp, and every generation is billed in Higgsfield credits, so the practical questions are less about setup and more about spending and content rules.

What the Higgsfield MCP server does

Higgsfield’s MCP page↗ (opens in a new tab) describes access to more than 30 image and video models from inside your assistant, plus a catalog of ready-made workflows for paid ads, marketing, motion design, production and short-form video. Its companion pages add the detail:

• Images and video from text, reference images, or both. The Higgsfield CLI page↗ (opens in a new tab) gives images up to 4K and videos up to 15 seconds, and says a connected agent can also train characters and browse your creation history.
• Model choice. Higgsfield lists its own models, such as Soul and Cinema Studio, next to third-party video and image models. Which ones you can use depends on your plan.
• Consistent characters. Higgsfield’s Soul characters keep a face or avatar consistent across shots, which matters when a campaign needs the same person in several clips.
• Workflows that reach local apps. Some presets hand the result to After Effects, Blender, Premiere Pro, Photoshop or DaVinci Resolve Studio, and Higgsfield notes that these need the agent on the same computer as the app.

Higgsfield does not publish a tool-by-tool reference on its own pages. Once you connect, your client lists the server’s tools by name, and that list is the one to read before you approve anything.

The remote MCP server URL

Higgsfield’s pages do not quite agree on the address. The MCP page’s copy button and connector link use https://mcp.higgsfield.ai/mcp. Higgsfield’s guide to generating video from Claude↗ (opens in a new tab), last updated August 29, 2026, gives https://mcp.higgsfield.ai without the path. Use the one on the MCP page; it is what Higgsfield’s own one-click Claude link sends. If a client refuses the shorter form, the missing /mcp is the reason.

Setup in Claude

1. In Claude on the web or the desktop app, open Customize, then Connectors, and choose Add custom connector.
2. Name it Higgsfield and paste https://mcp.higgsfield.ai/mcp.
3. Select Connect and sign in to your Higgsfield account in the window that opens.
4. Start a new chat and try a small request first: one short clip, not a batch.

Anthropic’s guide to custom connectors↗ (opens in a new tab) covers the plan details: on Team and Enterprise, only an Owner can add a custom connector, under Organization settings, before members can turn it on. The same page warns that you should connect only servers you trust and review Claude’s tool approval requests. Higgsfield’s guide suggests setting its tools to Always Allow for an uninterrupted workflow; keep approval on for generation tools until you know what each call costs.

Setup in ChatGPT and other clients

In ChatGPT, Higgsfield ships as a plugin: open its plugin page, install Higgsfield, and sign in. For any other agent that supports remote MCP servers with sign-in, Higgsfield’s instructions are to add a remote server named Higgsfield with the copied URL and complete sign-in, with the exact controls depending on the agent. How that sign-in works under the hood is in MCP OAuth, explained.

Claude Code: the CLI route or the MCP route

For Claude Code, Higgsfield’s own setup is a CLI plus agent skills rather than a raw server entry: install @higgsfield/cli from npm, run higgsfield auth login and finish sign-in in the browser, then add its skills with npx skills add higgsfield-ai/skills. The skills repository↗ (opens in a new tab) is MIT-licensed and covers generation, character training, product photos, brand kits and more. Higgsfield’s page also tells you to review the commands the agent proposes before running them.

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

If you would rather connect the remote server directly, Claude Code adds any remote MCP server with its standard claude mcp add command; run /mcp inside Claude Code afterward to sign in. Use one route or the other, not both, so the agent does not see two copies of the same tools.

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

How credits are spent

The MCP server has no separate billing. Higgsfield says its agent tools use the same credit system as the platform, and that each generation costs credits based on the model and the resolution. New accounts come with some free credits, and paid plans add volume, longer clips and fuller model access. The terms of use↗ (opens in a new tab), last updated July 26, 2026, add two rules worth knowing: add-on credits are valid for 90 days from purchase, and unused credits expire when you cancel your account.

• Ask for a draft first. A short, lower-resolution clip tells you whether the prompt works before you spend on a long 4K one.
• Watch batches and variants. A request such as “make 10 versions” is 10 generations, and an agent will happily run all of them.
• Ask the agent to check your balance before and after a session, and to stop and ask before any step it expects to be expensive.
• Your assistant’s own usage limits apply on top, because every tool call and result also counts against your Claude or ChatGPT plan.

Content rules

Generating through an assistant does not change what Higgsfield allows. Its terms require you to be at least 18, or the age of majority where you live if that is higher. If a real person’s face, name or voice is in your input, you need their consent and any releases the law requires. You may not impersonate anyone, depict minors in any inappropriate way, infringe someone’s intellectual property or privacy, or try to get around a content filter or safety system. Higgsfield can refuse to process content it judges objectionable, and you may not present output as human-made or strip its provenance signals.

On ownership, the terms say Higgsfield does not claim your inputs or outputs and does not restrict commercial use of outputs, but it may use your content to train its models unless you delete it or work under an Enterprise agreement. For client work, that last point is the one to raise before you upload anything confidential. The terms also leave you solely responsible for how you use, distribute or rely on the output, which includes any ad it ends up in.

Higgsfield, Canva or both

Higgsfield generates the footage and images; it is not a layout tool. If the job is a finished flyer, deck or social post built from templates and a brand kit, that is what the Canva MCP server is for, and the two can sit side by side in the same assistant.

Tracking video requests on a board

Video work arrives as requests and leaves as drafts, and a chat thread is a poor place to keep track of either. With fenbs connected as a second MCP server at https://fenbs.ai/api/mcp, an assistant can file each request as a task in To Do, with the brief, the model, the aspect ratio and a link to the Higgsfield generation in the note, and move it to In Progress when it starts generating. A rule such as “always ask before any generation above 1080p” can go on the Decisions and rules page, which every connected assistant reads first, and the history shows who changed each task.

Be clear about what fenbs lacks. It has no due dates and no settable assignee, so a launch date or an editor’s name goes into the note, and it does not store media: the video stays in Higgsfield, and the task holds the link.