Skip to content
Neural Minute

Tutorial

How to Use Higgsfield in Claude Code From the Terminal

By Maxime Aulne, Founder and Editor

Higgsfield's terminal route is a CLI package plus companion skills, not the MCP connector. Install @higgsfield/cli, run higgsfield auth login, add the skills.

What you need before you start

Three things. Claude Code installed and working in the shell you intend to use. A Higgsfield account, because the setup includes a login step that binds the terminal to that account. And a working npx, since the companion skills are installed through it.

The commands published on higgsfield.ai/mcp are the contract. The package is @higgsfield/cli. The login command is higgsfield auth login. The skills install is npx skills add higgsfield-ai/skills. Those three strings are the whole setup, and everything else on this page is about doing them in the right order and knowing what they do not cover.

One framing point saves time. This is not the connector. The connector is a remote server at https://mcp.higgsfield.ai/mcp that you attach inside a chat client. The CLI is a local package that talks to Higgsfield from your shell and gives Claude Code a set of slash commands to work with. Both appear on the same MCP page, which is why the two get conflated, and the confusion costs people a setup that half works.

How the job is done with Higgsfield

Install the package @higgsfield/cli with the package manager on the machine. The MCP page names the package rather than an install line for every manager, so use the global install form your manager expects for a scoped package.

Then run higgsfield auth login. This is the step that binds the shell to your account, and it is the one people skip when they are in a hurry. Nothing downstream works without it.

Then run npx skills add higgsfield-ai/skills. The skills are the companion layer. They are what turns a request into a slash command inside Claude Code, so this step is the difference between a terminal that can reach Higgsfield and a terminal that can run a named preset.

The presets cited on the MCP page are worth reading before you invent your own workflow, because each one is a packaged job rather than a single model call. /katana handles aura edits. /ugc-flow handles product review UGC. /marketing-adapt-video is the Ad Recreator. /marketing-research-to-ads is the Ad Strategist. /marketing-hook-variants is the Hook Multiplier. /vox-mixed-media produces editorial motion graphics. /stickman-explainer does the explainer format. Seven presets, seven different shapes of work.

That list also tells you when this route beats the connector. Take the CLI when the work belongs in a repo or a terminal session, when you want a repeatable named preset instead of a fresh conversation each time, or when you are wrapping a generation step inside a build you already script. Take the connector when you want a chat client with no local install, and when the person doing the work is not at a terminal. Neither route is the wrong one. They fit different desks.

One thing neither route carries is the plan’s unlimited models and the free generations included in a paid plan. Those stay on higgsfield.ai, and the limit note on the pricing page excludes MCP, CLI, Canvas and Supercomputer together. The CLI is a faster door into the metered tools, not a way around the meter.

On this page

What to check in the result

Confirm the login held. Run a second command in the same shell after higgsfield auth login and make sure the session is still recognised. A login that succeeds and then disappears usually means it landed in a different shell profile than the one Claude Code is using.

Confirm the skills installed. After npx skills add higgsfield-ai/skills, a preset name typed into Claude Code should resolve as a command. If the preset is treated as ordinary text, the skills layer is not loaded, and that is a different fault from a model being unavailable.

Confirm which one you are looking at when a preset fails. A command that does not resolve is a skills problem. A command that resolves and then returns an error is an account, entitlement or model problem.

Then check the output itself. A preset that runs should return a generation rather than a description of one. If you get a tidy paragraph explaining what the clip would look like, the preset was not executed.

Where it usually goes wrong

The most common failure is installing the CLI and expecting it to behave like the connector, then reporting that Claude does not see Higgsfield. Claude Code sees a CLI and a skills set. A chat client sees a remote server. They are not interchangeable and they are not meant to be.

The second is running the skills install before the login. The order in the published route is login first, skills second, and reversing it produces a skills layer with nothing to call.

The third is treating a preset as a model name. /marketing-adapt-video is a workflow, not a checkpoint, and asking it for something outside its shape will disappoint without anything being broken.

The fourth is expecting the unlimited lane. A large tier bought for unlimited usage does not make unlimited usage appear in the terminal, because the limit note keeps that on the site.

The fifth is assuming the full catalogue is present on day one. Access to new models can arrive progressively and is not guaranteed for every account at launch, so a gap in the list is not evidence of a bad install.

What the plan allows

The terminal runs on the same allowance as everything else. The grid read on 7 November 2026 is unchanged by the client you use. Free is $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. Ultra is $129 for 3,000, with larger tiers at $250 for 6,000 credits and $375 for 9,000. On the business side, Team is $69 for 1,000 credits and Scale is $169.01 for 2,500.

Prices on that grid exclude VAT and local taxes, which are calculated at payment. If you are sizing a tier against a workload, the Higgsfield plans page is the page to read, since the annual view sits next to the monthly one there.

What we did not do

No account was created and no CLI was installed for this page. The login and the skills install are described from the commands the MCP page publishes, not from a run that was watched, so no output text from either command is quoted. The seven presets were not executed. No video, image or explainer was produced, and no timing for any of it is given, because nothing 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, no CLI was installed 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.