Higgsfield MCP: Generate AI Images and Video from Claude Code

Problem: A modern paid-media test cycle needs five hooks, three aspect ratios, and a fresh thumbnail per ad set. Building that in a creative tool means twelve browser tabs and four logins. Building it in Claude Code with Higgsfield MCP means one prompt and a polling loop.

Quick Win: Add the Higgsfield connector once, then ask Claude to "generate three 9:16 product hero shots with the headline 'Sleep cooler tonight' on the bottle and four 15-second TikTok videos using Kling 3.0." The agent picks the right model, sends the job, polls until ready, and returns URLs.

Higgsfield shipped its official MCP server on April 30, 2026, exposing 30+ image and video generation models through a single hosted endpoint. For anyone running paid creative, this is the missing layer between the prompt in your head and the asset on your hard drive. This post covers what launched, how to wire it into Claude Code, which models are worth your credits, and how it fits alongside the official Meta MCP and the Shopify AI Toolkit.

What Higgsfield Just Launched

Higgsfield is the AI media company behind Soul (photorealism), the Cinema Studio physics simulator, and a video roster that has dominated creator feeds since late 2025. Higgsfield AI also owns the cultural moment around signature effects: bullet time, earth zoom, the transition library that taught half of TikTok how to cut a hook. Until April 30, all of that lived behind a web UI and a REST API.

The launch is a hosted MCP server at https://mcp.higgsfield.ai/mcp. It speaks the Model Context Protocol over HTTP. Auth runs through your existing Higgsfield account and existing plan credits transfer. No API key to provision, no developer app to register.

What it bundles in one connector:

• 15+ image models including Soul 2.0, Nano Banana Pro, GPT Image 2, Seedream 4.0, FLUX, Reve, Kling O1
• 16+ video models including Sora 2, Veo 3.1, Kling 3.0 (and 2.6, 2.5 Turbo, 2.1), Kling o1, Seedance 2.0, Wan 2.7, MiniMax Hailuo 02
• 9 preset video formats tuned for ecommerce: UGC, unboxing, product review, TV spot, and others
• Signature effect templates including the bullet time, earth zoom, and transition libraries that put Higgsfield AI on creator feeds in the first place
• Soul Characters for training a consistent character, model, or product across a campaign
• Native 4K image output, video up to 15 seconds, every common aspect ratio

The pitch: instead of switching between Higgsfield's web app, OpenAI's Sora portal, Google's Veo waitlist, and a Kling subscription, you point Claude at one MCP. The agent reads each model's strengths, picks one, manages credit cost, and waits for the job.

Why an MCP Beats Traditional Image and Video APIs

Image and video generation has a workflow problem. The prompt-to-asset round trip is long (seconds for images, minutes for video), cost varies by model, and the "right" model depends on what the asset is for. Text-on-product needs Nano Banana Pro. Editorial photorealism needs Soul 2.0. Object-permanence physics in a 15-second clip needs Sora 2. Most operators pay for three subscriptions and still pick the model by hand.

A full higgsfield mcp claude code loop closes that gap three ways.

The agent picks the model. Tell Claude what the asset needs to do (sell, demo, hook, retarget) and model descriptions in the MCP guide selection. You stop memorizing which generator owns which surface. This is the core promise of any well-designed ai image generation mcp: the model is implementation detail, the asset is the deliverable.

The agent polls async. Video is not a synchronous API. Submit a 4K Veo 3.1 job and it takes three to five minutes. The MCP exposes a wait_for_job pattern (community wrappers like jfikrat/higgsfield-mcp document higgsfield_wait_for_job and higgsfield_get_job tools). Claude submits, sleeps, checks, surfaces the URL when ready.

The agent manages credits. Each model has a different credit cost. The MCP exposes balance and per-model pricing so the agent can warn you before burning a plan on a maximum-quality experiment.

The result feels closer to talking to an art director than driving a generator: describe the spot in the funnel, Claude picks Soul for hero, Nano Banana Pro for packaging text, Kling 3.0 for the social cut, and you wake up to a folder of named files. The Higgsfield prompts you used to write by hand (long, model-specific, full of mood-board language) become inputs the agent assembles itself from the brief.

Connecting Higgsfield to Claude Code

The setup is shorter than any other MCP I configured this year because there is no API key, no OAuth scope to negotiate, no .env file to maintain.

1. Open settings. In Claude Code, open Settings -> Connectors. The same flow works in Claude Desktop and on claude.ai for non-Code clients.
2. Add a custom connector. Name: Higgsfield. URL: https://mcp.higgsfield.ai/mcp. That is the entire configuration.
3. Authenticate. Click "Connect." You are redirected to Higgsfield's auth. Sign in with the same account you use at higgsfield.ai. Approve the connector. Existing plan credits are now visible.
4. Verify. Run List my available Higgsfield models and show my current credit balance. You should see the model roster and a credit count.

A note for higgsfield mcp users running multiple MCP servers: 30+ tool definitions cost context. Pair Higgsfield with MCP Tool Search so model schemas load on demand. Without it, Higgsfield plus Meta plus Shopify can chew through 12K tokens before you type your first prompt. With it, the cost drops to a few hundred until a tool is actually called.

The Model Roster: 30+ Generators in One MCP

The roster is the killer feature. Below is a working operator's guide organized by the funnel surface the asset is going to. Higgsfield's own catalog pages (higgsfield.ai/ai-image and higgsfield.ai/ai-video) confirm strength descriptions, and the jfikrat/higgsfield-mcp wrapper lists the model IDs the API exposes.

Image Models

Nano Banana Pro owns text rendering. Headlines, packaging copy, label legibility, ad text, anything where a model typo means a reshoot. Native 4K, native inpainting. If your hook is "FREE SHIPPING" overlaid on the product, this is the model.

Soul 2.0 is photorealism. Editorial-grade hero shots, lifestyle, model-on-product. When you would have hired a photographer for a $400-a-day catalog day, Soul replaces it. Also powers Soul Characters (below).

GPT Image 2 matches Nano Banana Pro for text rendering at 4K and is sharper when the composition is more illustrative than photographic. Worth A/B testing the two for any ad where text and image must read together.

Seedream 4.0 is the all-rounder. When you do not have a strong opinion about photorealism versus illustration, Seedream gets you 80% of the way without burning premium credits.

FLUX is the open-source workhorse and the cheapest credit-per-image option. Fine for thumbnails, A/B variants, anything that is not the hero.

Reve, Kling O1, WAN 2.2 round out the catalog with specialty strengths. Use them when the four above do not match the brief.

Video Models

Sora 2 owns physics. Bottles being poured, fabric draping, motion where object permanence and weight matter. OpenAI's "deep world simulation" framing is not marketing fluff, it shows up in the gen.

Veo 3.1 is the cinematic choice: 4K native, broadcast-grade visual flows. Pick Veo when the spot needs to feel like an ad, not a clip.

Kling 3.0 is Higgsfield's house favorite. The catalog calls it "the new standard in photorealism with advanced motion complexity." First-frame and last-frame control, native sound, strongest character consistency across multi-shot sequences. For a 6-second TikTok hook, Kling 3.0 is the safe default.

Seedance 2.0 is ByteDance's first native audio-video model. Lip-sync, SFX, music in a single pass. UGC-style spokesperson shot? Seedance saves you a separate audio pipeline.

Kling 2.6, 2.5 Turbo, 2.1 are the budget tiers. Use 2.5 Turbo for fast iteration, then re-render the winner on Kling 3.0.

Wan 2.7 is the speed-versus-richness sweet spot: fast iteration without the budget-tier look. Use it for variant batches. Kling o1 is the reasoning-tier video model, built for multi-layered scenes with a narrative beat. MiniMax Hailuo 02 is strong on character animation. Image2Video turns any still into a 5-second motion clip, the cleanest way to apply a Higgsfield transition effect to existing brand photography.

Pragmatic heuristic for paid creative: rough hooks on Kling 2.5 Turbo, render the winner on Kling 3.0 or Veo 3.1, reserve Sora 2 for the one or two flagship spots where physics carry the message.

The Three Operator Workflows

The model list is interesting. The workflows are what move money.

1. Product Photography at Catalog Scale

40 SKUs need fresh hero images for a seasonal landing page. White seamless, three-quarter angle, cool natural light, label legible. In Claude Code:

For each SKU in this CSV (sku, name, color), generate a Nano Banana Pro
hero image at 4096x4096, white seamless, three-quarter angle, cool
window light camera left, label fully legible. Save URLs back to CSV.

Claude reads the file, submits 40 parallel jobs, polls each, writes URLs back. A run that takes a photo studio a full day finishes in 20-40 minutes of gen time and costs less than a single billable hour.

2. UGC Ad Video for Paid Tests

You run a weekly Meta creative cadence and the bottleneck is the four to six new hooks per ad set. UGC briefs that used to require a creator round trip now generate from the desk:

Generate 6 9:16 UGC videos using Seedance 2.0, each 8-12 seconds,
lip-synced spokesperson, hooks: "I tried this for 30 days",
"If you have trouble sleeping", "Don't buy a topper until you watch
this", "This replaced my Magnesium routine", "POV: it actually worked",
"My doctor said try this first". Casual home, warm lighting, product
close-up ending. Reference: /assets/hero.png.

Six hooks at three iterations is 18 variants in a 90 to 180 minute generation queue. Same volume from a UGC marketplace is $1,200 to $3,000 and seven to ten days of turnaround.

3. Soul Character Training for Brand Consistency

This is the workflow that justifies the MCP for a Shopify operator. Soul Characters trains a reusable face (model, mascot, CEO, recurring spokesperson) you pass as a parameter to every subsequent generation.

Create a Soul Character "Maya" from /refs/maya-1.jpg through
/refs/maya-5.jpg. Then generate 20 4:5 images: Maya holding the
product across kitchen morning, gym afternoon, beach golden hour,
home office, park bench.

Soul Characters costs 40 credits to set up (per the community wrapper). From then on, Maya is a parameter. A full campaign with the same face across hero, lifestyle, UGC, and email banner becomes one Claude session instead of a casting call. For DTC brands building a face for the brand, this is the highest-leverage feature in the entire MCP.

From Creative to Campaign: Wiring Higgsfield to Meta and Shopify

Generation is the easy part now. The hard part is which model for which spot in the funnel, how the asset gets named and tagged, and how it ends up on the right ad set or product page.

This is where the higgsfield mcp sits inside a bigger operator stack. With three MCPs in one Claude Code session:

1. Higgsfield MCP generates the creative.
2. Meta MCP ships it to ad sets and pulls performance back (install guide).
3. Shopify AI Toolkit drops assets onto PDPs, updates metafields, edits theme sections (setup guide).

A real session prompt: "Generate three Soul 2.0 hero variants for the Sleep Cool Bundle, upload to the Shopify product sleep-cool-bundle as primary image, then create a Meta Advantage+ creative test using all three with headline 'Cool down faster.'" Three MCPs, one prompt, one polling loop. Without the distribution rails, Higgsfield is a faster Photoshop. With them, it is the creative arm of an autonomous merchant.

The Shopify Kit's media-creation/ skills (7 files: product photography, lifestyle, UGC briefs, short-form video, asset organization) are written for exactly this. Feed Higgsfield your Kit brand-template variables (the _brand-template/ files ship pre-structured) and you get on-brand output without re-prompting every gen. The Kit's asset-organization playbook tells Claude where to save, how to name, which campaign tag to apply, so a 40-image run lands as a 40-row CSV your team can ship without renaming.

Pricing, Credits, and the Async Pattern

Higgsfield bills on a credit system. Free tier ships a daily allowance; paid plans range from creator-tier monthly up to unlimited at the top end. Per-generation cost varies by model: a Soul 2.0 1080p image is cheap, a 4K Sora 2 video is expensive, Soul Character training is a fixed 40 credits.

Three things matter for higgsfield mcp session economics:

1. Check credits before a batch. A 200-image catalog is fine on a creator plan. A 200-video Veo 3.1 batch burns through a month in an afternoon. Ask Claude to estimate first.
2. Async is safe to interrupt. Jobs persist server-side. If your session ends, the job continues. Resume in a fresh session, list jobs, grab URLs. That makes long Sora 2 and Veo 3.1 runs safe to fire and walk away from.
3. Failure modes are surfaced. Jobs return queued, in_progress, completed, failed, nsfw. The agent retries failed jobs with softer prompts and rewrites NSFW flags. Matters when running unattended generation on a dozen variants.

The honest cost framing: high-volume operators will exhaust a creator plan in week one. The math still works. Replacing a $4,000-a-month UGC retainer or a $6,000 catalog photo day with even the top Higgsfield plan is a 5-to-10x cost reduction on the same output.

Comparing Higgsfield to Standalone Sora, Veo, and Runway

A fair question: why use Higgsfield MCP instead of OpenAI direct for Sora, Google for Veo, or Runway for their own stack?

Higgsfield MCP wins on aggregation and one auth. Thirty-plus models, one connector, one credit pool, one polling pattern. The time saved on context switching is the entire ROI. You stop maintaining four developer accounts and treat "model" as a parameter on a function call.

Standalone Sora is the right choice when you only ever need Sora and want OpenAI's web UI features (storyboard, remix). Higgsfield exposes Sora 2 generation, not the full portal.

Standalone Veo is right when you are deep in Google's creative stack (YouTube assets, Workspace integration) and single-vendor billing simplifies things.

Runway is right when you need the full timeline editor, motion brushes, and green-screen workflow Runway built around its models. Higgsfield is generation-first; Runway is a creative tool with generation inside.

Higgsfield plus Claude Code is right when the asset is the deliverable, the deliverable is going to a Meta ad set or a Shopify PDP, and you would rather describe the brief than learn another timeline UI. For 90% of paid-creative and ecommerce operators, that is the working answer.

Closing the Loop

Two years ago, faking a believable product photo took three subscriptions, six tabs, and an inpainting workflow that broke every time a tool updated. Today, you point Claude at one MCP endpoint and the shot lands in the asset folder before your coffee is cold.

The new part is not the models. Sora and Veo have been impressive for months. The new part is the integration: those models now sit behind the same agent that runs your ads and edits your store. The asset is no longer the bottleneck. Judgment is, which model for which surface, which hook for which audience, which version goes to the cold ad set and which goes to retargeting.

That last layer of judgment is where it gets operator-specific. The Shopify Kit ($199) ships an asset-organization system and 7 media-creation playbooks (product photography, lifestyle, UGC briefs, short-form video, asset organization, brand-template variables, named-file conventions) for exactly that. Pair it with Higgsfield MCP and you have a creative pipeline that runs from your terminal: prompts matched to your brand, output landing in the right folder, clean handoff into the Meta MCP and Shopify AI Toolkit so the asset reaches the customer without three more tools in between.

The MCP is free to connect. The infrastructure to use it well is what changes the math.

Next Steps

• New to MCP? Start with MCP fundamentals before adding 30+ generation tools.
• Running multiple MCPs? MCP Tool Search drops Higgsfield's context cost from 12K tokens at startup to a few hundred until a tool is called.
• Browse the curated MCP list for the rest of the operator stack.
• Compare Higgsfield against the social-media MCP landscape, where most servers can post but none can create.
• Pair Higgsfield with the Meta MCP and the Shopify AI Toolkit for the full creative-to-checkout loop.