Create Images and Videos with an AI Agent: Higgsfield MCP in Claude Code

Higgsfield connects to Claude Code with a single claude mcp add command and the address https://mcp.higgsfield.ai/mcp. After that, you give it a task in plain language, and the agent goes to Higgsfield on its own, generates an image or video, and returns it right in the window. You pay twice for this: with Higgsfield credits for generations and Claude Code limits for having the server in context. So my rule is simple: if you aren't using MCP, turn it off.

The old process was long: go to the Higgsfield website, figure out the interface, write a prompt, generate, download, and upload the result to your project. Now Higgsfield has released its own MCP, so you don't have to go to Higgsfield or figure out those interfaces at all.

It's hard to imagine anyone going into Higgsfield to generate something now, when you can generate everything through your own Claude Code.

I've made an entire presentation with this setup, including all the videos in it. I described the whole task and handed it to the agent.

One danger: the agent might go and churn out 100 images, and you pay for every one. So include a condition about cost and the number of attempts in every task.

Let's go through it in order: how to connect the server, how to give it a first task, what you pay with, and when to turn the server off.

Step 1. Add the Higgsfield server with the claude mcp add command#

Open a terminal in the project folder where you'll generate images and videos, and run:

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

The address is on the official higgsfield.ai/mcp page; copy it with the Copy connector URL button.

Without the --scope flag, the server is added only to the current project. With --scope project, it will be available to the whole team on that project; with --scope user, it will be available in all your projects.

Add Higgsfield where you work with it. In a project where you write code or text, it will take up context space without doing anything.

I covered other servers for Claude Code in a separate article.

Step 2. Sign in to your Higgsfield account through /mcp#

Launch Claude Code and enter /mcp. Select higgsfield from the list; a browser will open. Sign in to your Higgsfield account.

Generations will be charged to this account. Until you sign in, the server returns a 401 error: Claude Code can see it but can't use it.

Step 3. Check that the server connected#

Run this in the terminal:

claude mcp list

The higgsfield server should show the status ✔ Connected.

If the status is different, check the details: the address, scope, and server status.

claude mcp get higgsfield

If that doesn't help, remove the server and add it again:

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

Afterward, sign in again through /mcp. Claude Code also erases the saved login when it removes the server, so you'll get the same 401 error without signing in again.

How can I set a cost boundary for Higgsfield MCP generation?#

Give it a first task without generating anything. Have the agent show you what the server can do before you spend anything:

Show me which tools and models the Higgsfield server has and how many credits generation costs with each model. Don't generate anything.

Next, give it the actual task. Add one condition to the end of each task: before paid generation, show the cost and agree on the number of attempts.

A /goal prompt with an attempt limit works on the same principle: the agent has a limited number of attempts in advance.

Task for an image for the hero section of a website:

Prepare an image in Higgsfield for the hero section of my product's website. Product: [link]. Suggest an object, angle, color, and setting that help explain the product. Leave room in the frame for the page heading and button. Horizontal format. Save the finished file in the project's images folder.
Before paid generation, show me the cost, agree on the number of attempts, and wait for my “yes.”

Task for a video made from this image:

Make a short video in Higgsfield from the image images/hero.png: the camera slowly moves toward the product, and the background moves slightly. Suggest a model and ask me how long the video should be. Make 1 version.
Before paid generation, show me the cost, agree on the number of attempts, and wait for my “yes.”

The agent then goes to Higgsfield on its own, writes a prompt, generates, gets an image or video, and delivers it to you in the Claude Code window. Tell it what to change in plain language too.

Which models are available and what you pay for generations#

Higgsfield works well for me because it has all the AI tools built in, included with the subscription, although you pay with credits.

Ask the agent for the model list with the first task in the previous section; you'll get the models and their credit prices as of today.

Generations through MCP are charged to your Higgsfield account in credits. The server itself uses Claude Code limits, so you have two usage meters. I showed where Claude Code tokens go in a separate article:

How much usage the MCP server consumes and when to turn it off#

Every MCP server uses limits: descriptions of its tools are added to the model's context.

Tool search is enabled by default in Claude Code. At the start of a session, the server's tool names and instructions are loaded; full descriptions are loaded when needed. If Claude Code runs through your own ANTHROPIC_BASE_URL, search is disabled and the descriptions are loaded in full. In both cases, the server takes up space in the context of every session.

If you have a huge number of MCP tools installed, they eat up your limits without doing you any good. So here's an important rule: if you aren't using MCP, turn it off so it doesn't overload the context window.

You can turn the server off and back on like this:

• Temporarily. Use the toggle next to higgsfield in /mcp. Claude Code stops connecting to it, the setting is saved, and the selection is remembered for this project.
• Completely. Run claude mcp remove higgsfield. The saved account login is erased along with the server.
• Turn it back on. Use the same toggle in /mcp, or run claude mcp add again and sign in through /mcp.

Keep the server on while you're working with images and videos. If you're writing code or text and have nothing to generate, turn it off.

Which approach to choose: MCP, the Higgsfield program, or an unofficial server#

There are three ways to connect Higgsfield to an agent:

Start with the MCP server: it takes one command to install and one toggle to turn off. If you generate every day and don't want a server taking up context, look at the Higgsfield program with skills.

Next step: In the connected Higgsfield account, run the no-generation tools, models, and credit-cost check above. Choose one intended output and an attempt budget before any generation. For a separate Ailoot API workflow, follow the image request guide or the video task and result guide; those describe Ailoot API calls rather than this Higgsfield MCP account.