Skip to content
Neural Minute

Tutorial

How to Connect the Higgsfield MCP Server to Claude

By Maxime Aulne, Founder and Editor

Higgsfield runs one MCP server at https://mcp.higgsfield.ai/mcp and Claude adds it from a pre-filled screen. Plan credits stay on the web, not the connector.

What you need before you start

Higgsfield runs one MCP server, and the address is https://mcp.higgsfield.ai/mcp. That single URL is what every client on the list points at, whether the client is ChatGPT, Claude, Claude Code, Telegram, the CLI, or an entry filed under Other.

You need three things. A Higgsfield account. A client that can attach a remote MCP server. And a clear idea of which of your Higgsfield entitlements the connector actually carries, because that is where most people lose an afternoon. Unlimited models and the free generations included in a paid plan are available on higgsfield.ai only. They do not travel through MCP, and they do not travel through the CLI, Canvas or Supercomputer either. The connector gives you the generation tools. It does not give you the plan’s unlimited lane.

Nothing else is required. No separate MCP subscription is listed, no API key is asked for on the MCP page, and the connector is not sold as an add-on to the grid.

How the job is done with Higgsfield

Open higgsfield.ai/mcp and pick your client from the list on that page.

For Claude, the page carries a direct link into Claude’s add-connector screen. The link opens the form with two fields already filled, the name Higgsfield and the address https://mcp.higgsfield.ai/mcp. You confirm, and the server is attached to the account. That pre-filled route is worth using even if you know how to add a connector by hand, because a hand-typed URL is the first cause of a connector that appears in the list and never answers.

For ChatGPT, Telegram, Claude Code and the CLI, the same page routes you to the instructions written for that client. The Other entry covers anything not named, which in practice means any client that accepts a remote MCP endpoint and lets you paste a URL.

Once the server is attached, the work happens in the chat. You describe the clip or the image you want, the client calls the Higgsfield tool, and the model runs on Higgsfield’s side. You do not write request payloads by hand on this route. If you want payload-level control and a price per second, that is the API on open.higgsfield.ai, which is a different product with a different bill.

The terminal is on the same page and is a separate path. The package is @higgsfield/cli, the login command is higgsfield auth login, and npx skills add higgsfield-ai/skills installs the companion skills that turn a request into a slash command such as /ugc-flow or /marketing-hook-variants. Treat that as its own setup rather than as a step inside the connector flow. Adding the remote server and installing a local package solve different problems, and doing half of one does not improve the other.

On this page

What to check in the result

Check the connector before you judge any output. The entry should read Higgsfield, and the URL should end in /mcp with no trailing space and no shortened form. A connector that shows a name but points somewhere else will fail in a way that looks like a model problem.

Then check that the client is actually calling the tool. A model with the connector attached but no instruction to use it will happily write a paragraph describing the video you wanted. The tell is a response made of prose about a generation rather than a reference to a generation. Ask for one clip at a time, not a storyboard, and the difference becomes obvious on the first call.

Check the model name that comes back. Higgsfield’s catalogue is broad and access to new models can arrive progressively, so a named model missing from your account at launch does not mean the connector is broken.

Finally, read the error when one appears. An authentication failure and a rejected request look different and need different fixes. Re-authenticating will not repair a request the model refused, and rewriting the prompt will not repair a missing session.

Where it usually goes wrong

The first mistake is expecting the plan’s unlimited models inside Claude. They are not there. Higgsfield’s own limit note puts unlimited models and the free generations of a plan on higgsfield.ai, and excludes MCP, CLI, Canvas and Supercomputer. If you bought a large tier for the unlimited lane and then moved your work into a chat client, you moved out of the lane you paid for.

The second is treating the connector as a replacement for the web app. It is a second door to the generation tools, not a mirror of the interface, and the parts of the product that are not generation tools do not follow you through it.

The third is mixing the two setups. Pasting the MCP URL does not install the CLI, and installing the CLI does not attach the server to ChatGPT.

The fourth is reading an unlimited allowance as unlimited speed. Higgsfield notes that unlimited usage can see dynamic speed adjustments when demand is high, so a busy hour can slow the lane without changing what the plan includes.

What the plan allows

The connector does not change your tier. The grid is the one the pricing page listed on 7 November 2026. Free costs $0 with limited daily credits and a selection of models and features. Starter is $19 a month for 270 credits. Plus is $59 for 1,200 credits. Ultra is $129 for 3,000 credits, with larger Ultra tiers at $250 for 6,000 and $375 for 9,000. The business tiers are Team at $69 for 1,000 credits and Scale at $169.01 for 2,500 credits.

So the connector is a convenience layer over the allowance you already hold, and the allowance is what limits the work. If you are choosing a tier before wiring anything up, the Higgsfield plans page carries the current grid and the annual view in one place.

What we did not do

No account was created for this page and no client was connected. The connector screen is described from the link the MCP page publishes, not from a completed setup, so the exact wording of the confirmation step is not reproduced. The Telegram handshake, the CLI login and the skills install were not run. No video or image was generated through the connector, and no render time appears anywhere in this silo because none was measured.

Sources

  • Higgsfield MCP page, higgsfield.ai/mcp, read on 16 November 2026.
  • Higgsfield pricing page, higgsfield.ai/pricing, read on 7 November 2026.

The plans, the credit conversions and the current campaign sit on the Higgsfield price page, and the full review is at /reviews/higgsfield/. ## Update log

16 November 2026. First version written from the MCP page and the pricing page. No account was used and nothing was generated.

Newsletter

One email when a review lands.

That is when the promo codes are still fresh, before the prices move.

Nothing else. Unsubscribe in one click.