How to Use Higgsfield in Claude, Cursor and ChatGPT (2026)

Higgsfield now runs inside the AI assistants you already use. You can connect it to Claude as a connector, add it to Cursor as a plugin, install it in ChatGPT as a plugin, or give Claude Code a command-line tool and a set of skills. Then you ask for an image or a video in chat and the assistant calls Higgsfield for you. Higgsfield's own guide to this, "How to Generate AI Videos With AI Agents in 2026", went up on September 28, 2026.

This post walks through the setup for each assistant, what each one can and cannot do, and how the credits work, which is where most people get surprised. One note on sources: we have not run these connections ourselves. This is a check of what Higgsfield documents, not a tested walkthrough. Everything below comes from Higgsfield's help center, blog, CLI and Cursor pages and its GitHub repos, checked on October 1, 2026, and where Higgsfield's own pages disagree we say so.

If your real question is whether Claude can make video at all, and by which routes, start with Can Claude make videos? instead. This guide is only about the Higgsfield route.

Which connection each assistant uses

Higgsfield uses a different mechanism per assistant:

All of them sign in with your Higgsfield account through the browser. You do not need an API key for any of these. That is different from the Higgsfield API, a separate product covered further down.

Set up Higgsfield in Claude

The MCP server address is https://mcp.higgsfield.ai/mcp.

1. Open Settings, then Connectors. (Higgsfield's blog words the path as Customize, then Connectors; it is the same place.)
2. Click Add custom connector.
3. Name it Higgsfield and paste the MCP URL.
4. Click Connect and approve your Higgsfield account in the window that opens.

Then ask in plain language. According to Higgsfield, Claude can reach 30+ models this way, including Soul, Cinema Studio, Seedance 2.5, Kling, Veo, Minimax Hailuo, Flux, Seedream and Nano Banana Pro. Claude can pick the model, or you can name one. Beyond images and video, the Claude connection exposes upscaling, reframing, background removal, motion controls, audio (voiceover, voice cloning, voice change, dubbing), multi-step "production skills" and your Higgsfield generation history.

Set up Higgsfield in Claude Code

Higgsfield's instructions for Claude Code are not the connector. They use a command-line tool plus skills, three commands in your terminal:

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

higgsfield auth login opens a browser sign-in. If you prefer Claude Code's plugin marketplace, the skills repo documents this instead (you still need the CLI installed and logged in):

/plugin marketplace add higgsfield-ai/skills
/plugin install higgsfield@higgsfield

The skills repo (MIT license, version 0.13.0 at the time of writing) ships 8 core skills, including higgsfield-generate, higgsfield-soul-id, higgsfield-product-photoshoot, higgsfield-video-explainer and higgsfield-websites.

To check the install, the repo's INSTALL.md says to ask your agent: "Generate a minimal test image with Higgsfield." It should call the higgsfield-generate skill and run higgsfield generate create z_image --prompt "test" --wait. Like any generation, this test spends credits.

Some third-party guides instead add Higgsfield's MCP server to Claude Code directly with Claude Code's own claude mcp add command. Higgsfield's own pages do not give that exact command; its Claude Code instructions focus on the CLI and skills.

Set up Higgsfield in Cursor

1. Open Customize, then Marketplace.
2. Find Higgsfield and click Add.
3. Sign in, then ask for an image or a video.

Higgsfield's Cursor page describes the same thing as the Plugins menu with a Connect button; the menu name may differ by Cursor version. You can also type /add-plugin in Cursor chat and search for Higgsfield. The plugin's source is public at github.com/higgsfield-ai/cursor-plugin under the MIT license.

Higgsfield's Cursor page makes two useful claims. First, only your generation prompt is sent to Higgsfield, not your code, files or workspace context, unless you include them in the prompt yourself. Second, most outputs render in 20 to 40 seconds, depending on length and complexity.

Set up Higgsfield in ChatGPT

ChatGPT does not use the MCP connector. Open the Plugins Directory (or higgsfield.ai/mcp), find Higgsfield, click Add, then sign in and approve access. Higgsfield notes that plugin availability can depend on your ChatGPT plan, region and workspace settings, without naming which plans.

The main limit: audio generation and the Website Building skill are not available in ChatGPT. Images and video work in the Claude connector, the ChatGPT plugin, the Cursor plugin and the CLI.

Codex: listed, with an open issue

The skills repo lists Codex alongside Claude Code and Cursor, installed with npx skills add higgsfield-ai/skills. Higgsfield's AI-agents blog post does not mention Codex.

An open issue on the CLI repo (#80, opened September 4, 2026) reports that the skills show up in Codex and can be selected, but Codex's skill checker flags two settings in the skills' files, argument-hint and version, so they do not pass its validation without local edits. A Higgsfield collaborator replied "checking" on September 14, and the issue was still open when we looked on October 1. The report tested skills version 0.12.0, so if you are on Codex, check the issue to see whether the current version fixes it.

What it costs

This is the part to read twice.

• A paid Higgsfield subscription is required. The help center says an active paid subscription is required; Higgsfield's Claude MCP blog post adds that there is no extra fee for using MCP on top of it.
• Every agent generation spends credits. The help center's wording: "Every generation through a connected agent deducts credits, regardless of your plan." Generations through MCP, the CLI and the ChatGPT plugin come out of your normal balance at standard rates.
• Unlimited and free generations apply only on higgsfield.ai. If your plan includes unlimited use of a model on the website, that does not carry over to Claude or Cursor.
• There is no built-in credit cap. Put a budget in your prompt ("one draft clip, tell me the cost before you make more").

You may see an "Unlimited MCP" offer in older posts. That was a 24-hour trial for new users, available until July 31, 2026. It is over.

How many credits a video takes, per Higgsfield's own examples: its ChatGPT guide lists a 15-second UGC video on Seedance 2.5 at 1080p as 270 credits, a 30-second cinematic take on Seedance 2.5 as 540 credits, and a character sheet on Soul 2.0 at 2K as 0.5 credits. Its marketers post prices a 20-second UGC product video at 270 credits ($13.50), so the two posts disagree on how long a 270-credit video is.

Plan prices: we couldn't confirm current plan prices or how many credits each plan includes, so check higgsfield.ai/pricing before you subscribe.

The Higgsfield API is a different product

The Higgsfield API (docs.higgsfield.ai, keys from console.higgsfield.ai) is not how the assistants connect. It is pay-as-you-go from a prepaid US dollar balance with no subscription, video costs about $0.0738 to $0.25 a second, and plan credits and Unlimited do not apply to it. Failed or moderated requests are not charged, and there are Python and TypeScript SDKs. Use it if you are building your own app; use the connector or CLI if you just want to generate from chat.

What you can and cannot do this way

You can:

• Generate images up to 4K and video in 16:9, 9:16, 1:1 or 4:5.
• Let the assistant choose among 30+ models, or name one.
• Use audio and voice tools in Claude, Cursor and the CLI (not ChatGPT).
• Find everything you made in your Higgsfield account afterwards, tagged with MCP as the source.

What to plan around:

• Clip length depends on the model and workflow. Higgsfield's CLI and Claude pages say up to 15 seconds per clip, but its ChatGPT guide shows a single 30-second Seedance 2.5 take. Check the length before you plan a longer piece, which may mean several generations that you or your agent edit together.
• You are still the director. The assistant writes the prompts, but deciding what the video should be, judging each clip and choosing what to keep is your job.
• Spending is open-ended unless you set a limit in the conversation.

A different job in the same assistants: make your version of a winning video

Higgsfield in your assistant gives you 30+ models and multi-step production skills, and you can steer it toward almost any shot you can describe. Riffkit's skill is built for one specific job: you point it at a short video that already won, and it makes your own version for your product.

The Riffkit skill runs in Claude Code, Codex and Cursor. You say one sentence, like "make my version of this winning video for my product", and the agent calls Riffkit's API. You start from a template in its library, a TikTok link or an upload (up to 45 seconds). Riffkit studies why the video worked (the hook, the pacing, the beats) and generates brand-new footage; it never re-uploads the source clip. Two modes fit this job:

• Adapt keeps the formula and tells a new story for your product.
• Swap keeps the shots, camera and action, and changes the person, product or details.

You can run one template with several characters in one batch, reframe a finished video to 9:16, 1:1 or 4:5 (re-rendered, not cropped), and generate natively in 9 languages. It renders on Seedance 2.0, Seedance 2.0 Fast, Seedance 2.5 or MiniMax H3, billed per second of finished video, and failed renders are not charged. On a monthly plan a 15-second video costs about $2.70 to $3.00 on MiniMax H3, depending on the plan. New accounts get free credits with no card. It does not connect to ad accounts, launch ads or read ad performance, and it does not make static image ads or carousels.

The day-to-day workflow is in Claude Code video skill, and if you would rather skip the agent, the same thing runs in the browser at riffkit.ai/app. For Higgsfield's reshoot feature specifically, see how to use Higgsfield Genjutsu.

Which one to use

• If you want to describe a shot and get it, with a choice of 30+ models, images up to 4K and voice tools, use Higgsfield in Claude, Cursor or Claude Code.
• If you already pay for Higgsfield and mostly work on the website, keep heavy use there: unlimited and free generations only apply on higgsfield.ai, while agent generations always spend credits.
• If you work in ChatGPT and need audio, use Claude, Cursor or the CLI instead; ChatGPT's plugin has no audio.
• If you are on Codex, check issue #80 on higgsfield-ai/cli first.
• If you are building your own app, look at the Higgsfield API, which bills a prepaid dollar balance with no subscription.
• If you start from a short video that already won and want your own version of it, with your product, person or language, use Riffkit's skill in the same assistant or the web app.

Both can live in one agent: use Higgsfield when you are building from a description or a brief, and Riffkit when your starting point is a video that already worked.