AI product photography
Higgsfield's new MCP connector lets you run its image and video models directly inside Claude, and paired with a free Claude Cowork project template, it turns one product photo into multiple professional product images and a full commercial in a single local workflow — but Thomas Lundström argues it works best when used as a step-by-step creative assistant rather than left on full auto.
Higgsfield's MCP connector lets you use Higgsfield's models directly inside Claude (desktop or browser), removing the need to copy-paste prompts between separate platforms.
Installing the connector involves copying a URL from Higgsfield's MCP and CLI page and adding it as a custom connector in Claude desktop's Connectors settings, then signing in with your Higgsfield account.
Connector permissions can be set to 'always allow' or to require Claude to ask before generating images/videos, as a way to manage credit spending.
Claude skills such as the image prompt helper skill and the Seedance 2 multiprompter (both free downloads) now work directly inside Claude alongside the Higgsfield MCP, speeding up asset creation.
Claude Cowork 'Projects' are folders on your computer that Claude can read, restructure, and act on according to custom instructions given for that folder.
The presenter provides a free downloadable project template (a CLAUDE.md file, a README file, and a RUN file) that automates turning one product image into multiple professional photos and a commercial.
Setup requires enabling Network Egress in Claude's Capabilities settings and adding two specific domains to the allowlist so Claude can upload and download files directly to and from the project folder.
The presenter recommends keeping the domain allowlist restricted to 'package managers only' plus the two added domains, rather than opening access broadly, citing security risk.
Running the workflow: drop a product image into the folder, tell Claude to "run it," and it creates a subfolder, optionally asks for style references, then asks about commercial style, environment, lighting, use case, and platform.
Style references can be either copied directly into the outputs or used only for their general 'vibe.'
A full-auto option exists, but the presenter strongly recommends the step-by-step assistant mode so the user retains creative control over outputs.
Claude proposes short frame descriptions (not the final prompts) before generating images, and the user can reroll or revise them.
Once frame descriptions are approved, Claude generates the images via the Higgsfield MCP, and the user can chat back and forth with it to refine specific images.
The next stage builds a multi-shot prompt for Seedance 2 from the created images, asking for the number of shots, editing style, and shot order; Seedance fills in gaps if more shots are requested than there are frames.
The final multi-shot prompt is a long text description of every shot, which can be edited or run as-is, and the workflow ends with a generated product commercial.
A drawback noted is that using Higgsfield inside Claude consumes Claude subscription tokens, which can add up quickly with complicated workflows.
Higgsfield MCP connector — A Model Context Protocol connector that lets you use Higgsfield's image and video generation models directly inside Claude (desktop or browser) instead of switching between separate platforms. Apply: Go to Higgsfield's MCP and CLI section, select the MCP tab with Claude chosen, copy the URL, then in Claude desktop go to Customize > Connectors > Add custom connector, name it 'Higgsfield', paste the URL, and connect/sign in with your Higgsfield account.
Connector permission settings (always allow vs. ask first) — A per-action permission setting in Claude that controls whether Claude generates images/videos automatically or asks for confirmation first, used to manage credit spending. Apply: In the connector settings, set actions to 'always allow' for speed, or require Claude to ask before generating images and videos if you want tighter control over credit usage.
Image Prompt Helper skill — A free downloadable Claude skill that helps generate image prompts, usable directly inside Claude once the Higgsfield MCP is connected. Apply: Download the skill from the video description and use it within a Claude chat connected to the Higgsfield MCP to speed up image prompt creation.
Seedance 2 Multiprompter skill — A free downloadable Claude skill that builds multi-shot prompts for the Seedance 2 video model. Apply: Install the skill and let Claude use it to generate a multi-shot commercial prompt from a set of already-created product images.
Claude Cowork Projects — A Claude Cowork feature where a project is a folder on your computer that Claude has direct access to, letting it read, restructure, and act on files per custom instructions for that folder. Apply: In Claude desktop, go to Cowork > Projects > New project > Start from scratch, name it, choose where it should live on your computer, and hit create.
Project workflow files (CLAUDE.md, README, RUN) — Three files provided as a free downloadable template — a CLAUDE.md instructions file, a README explaining the pipeline, and a RUN file that is the workflow itself — that together define the automated product-visual pipeline. Apply: Download the workflow files from the description and add them into your newly created project folder before running the pipeline.
Network egress domain allowlist — A Claude Capabilities setting that must be enabled to let Claude upload and download files directly from and to your project folder via specific external domains. Apply: In Claude's settings panel, go to Capabilities, enable Network egress, keep it restricted to 'package managers only,' and separately add the two required Higgsfield-related domains to the allow list rather than opening access to any domain.
"Run it" pipeline trigger — The workflow's activation step where dropping a product image into the project folder and telling Claude to "run it" kicks off the automated creation pipeline. Apply: Add a product reference image to the project folder, open the Cowork project, and type "run it" to start the sequence of subfolder creation, style-reference prompts, and generation questions.
Style references (copy vs. vibe) — An optional input where additional reference images influence the generated outputs, with a choice of copying them directly or only capturing their general 'vibe.'. Apply: Add style reference images to the project folder when Claude asks for them, then choose whether the workflow should copy the references directly or just use them as loose stylistic inspiration.
Full auto vs. step-by-step assistant mode — Two modes for running the pipeline: a full-auto option that lets Claude generate everything with minimal input, versus a recommended step-by-step mode where the user reviews and directs each stage. Apply: Choose full auto only if you want to 'go full bananas' with minimal oversight; otherwise use the step-by-step mode so you retain creative control and avoid wasting credits on unwanted results.
Frame description review/reroll — An intermediate step where Claude shows short descriptions of the planned image frames (not the actual prompts) before generating anything, so the user can catch mismatches early. Apply: Review the frame descriptions Claude proposes and ask it to reroll or revise any that don't match your intent before telling it to run and generate the actual images.
Multi-shot prompt generation for Seedance 2 — A step where Claude compiles the approved product images into a single long multi-shot text prompt for the Seedance 2 model, describing every shot of the eventual commercial. Apply: After finalizing your images, tell Claude to proceed to the commercial step; choose the number of shots (Seedance fills gaps between frames if you request more shots than frames), pick an editing style and frame order, then review or edit the resulting multi-shot prompt before letting it run.
The presenter frames 'use it as an assistant, not full autopilot' as the single most important, under-discussed point of the whole workflow, directly contrasting it with hype elsewhere ('everyone on Instagram and everywhere is just touting how Claude can automate everything...and it will create 500 ads for you') to argue manual, step-by-step creative control yields better results and conserves credits.
Restricting the network-egress domain allowlist to 'package managers only' plus explicitly whitelisted domains, instead of open access, is framed as a deliberate security safeguard against giving Claude unrestricted download capability.
The pipeline running entirely locally — Claude uploading and downloading files directly to and from a folder on the user's own computer — is presented as the standout advantage that outweighs the token-cost drawback of routing Higgsfield through Claude.
The frame-description review step (short text summaries shown before actual image generation) functions as a low-cost checkpoint to catch mismatched creative direction before spending credits on generation.
«Today, we're testing out the new Higgsfield MCP connector right inside of Claude.»
— 00:00
«There's no more jumping between workflows where you have to create a prompt in one place and then copy and paste that to another place.»
— 00:27
«I think it's important that you keep it on package managers only and then allow these domains separately because giving Claude access to any domain out there with download capabilities might be a security risk for your system.»
— 04:16
«And here we come to the most important part in my opinion that nobody's talking about because everyone on Instagram and everywhere is just touting how Claude can automate everything and you can just go to sleep and it will create 500 ads for you, but I would strongly recommend to use it more like an assistant if you care about your credits because if you go full auto, then it might produce something nice, but it's always better to have the creative control and freedom yourself.»
— 05:30
«Using Higgsfield inside of Claude, of course, uses some of the Claude tokens from your subscription. And especially if you build out some complicated workflows, it might use a lot of the tokens quite quickly.»
— 09:50
«It really felt like I didn't have to think about the prompting myself when I used it in this way because I can just chat with Claude that then uses the skills in the background to create the best possible results for me.»
— 10:16
Reception
Audience appreciates the tutorial with predominantly positive feedback and successful implementations, though some users find it too technical or express skepticism about the approach.
This is a sponsored, tool-specific tutorial (Higgsfield disclosed as sponsor) that maps a concrete, replicable Claude+MCP workflow with downloadable templates, useful mainly to viewers already working with Claude and Higgsfield for AI product visuals, and its central caveats — token cost and the case against full-auto — are self-reported by the presenter rather than independently verified.

10:48