Published: 2026-05-05

Claude Code + Higgsfield MCP Replaced a $5,000/Month Marketing Agency

Chapters / key moments (click to jump — plays here on the page)

Craig Hewitt cancelled a $5,000/month marketing agency after installing Higgsfield's MCP server into Claude Code — a 45-second setup that creates a CMO agent capable of building full marketing plans and generating creative assets on camera. The demo walks through adding the Higgsfield MCP connector in Claude Code's settings, then running a CMO agent skill that produces ICPs, channel strategies, and visual assets for a product launch. The framing: Claude Code is the brain, Higgsfield is the creative arms — together they function as a generative media agency inside a single chat window.

Source video

"Claude Code Just Saved Me $60K a Year (Here's How)" by Craig Hewitt — Watch on YouTube →

Key Takeaways

Higgsfield MCP installs in ~45 seconds: go to Claude Code settings → Connectors → Custom Connector, paste the Higgsfield MCP URL, and authenticate with your Higgsfield account via browser OAuth.
Higgsfield functions like an OpenRouter for generative media — a single MCP server that provides access to multiple image, video, and creative AI models (Higgsfield Soul, Cinema, GPT Image, and others).
The CMO agent skill works by reading a brand brief, producing an ICP analysis, channel strategy, and creative brief, then calling Higgsfield to generate the actual visual assets.
Compatible with Claude Code, OpenClaw, Hermes, NemoClaw, and Perplexity — any agent platform that supports MCP servers.
The full skills repo (CMO agent, brand brief template, and three demo scenarios) was shared for free alongside the video.
The model: Claude Code handles reasoning and planning, Higgsfield handles asset generation — you don't need to switch between tools or manually hand off work.

How the CMO Agent Skill Works

The CMO agent is a skill definition inside a local Claude Code repo. When invoked, it reads a brand brief file you've prepared — covering your ideal customer, marketing channels, product details, and brand voice. It uses that brief to produce a structured marketing plan: ICP definition, channel priorities, creative brief for each channel, and a production schedule. Then it calls the Higgsfield MCP server to generate the actual creative assets: images, product shots, or video clips using whatever generative models are configured. The output lands in a local marketing repo folder alongside the plan document. For direct-to-consumer brands, info product sellers, or SaaS companies, the setup requires only that you maintain an up-to-date brand brief file.

What Higgsfield Actually Is

Higgsfield is described as the generative media equivalent of OpenRouter — rather than routing requests to language models, it routes requests to creative and generative media tools. Through a single MCP server connection, you get access to Higgsfield's own image models (Soul, Soul Cinema), along with integrations to third-party generative tools. The practical effect is that Claude Code can generate a marketing plan and produce the visual assets for it without you logging into a separate image generation tool, downloading outputs, and re-uploading them elsewhere. The whole pipeline runs inside the Claude Code session.

Related on OpenClawDatabase

Claude Cowork Setup Guide — getting started with Claude Code and MCP connectors
Claude Cowork Skills Guide — building and running custom agent skills
Use Cases — real-world agent setups for business automation
OpenClaw Skills Guide — MCP and skills on the OpenClaw platform

What you can actually set up from this

Extracted from the video's own transcript — the specifics the original summary left out.

Reproducible steps

Install the media service as a custom MCP connector
In the desktop app: <strong>Code → settings (bottom left) → connectors → custom connector</strong>, name it, and paste the MCP URL from the service. Connecting opens a browser to authenticate against your account. You can then scope which tools and models it exposes.
Separate the brain from the arms
The framing the whole workflow rests on: <strong>Claude Code is the brain, the media service is the creative arms</strong>. One does the planning and the prompt writing; the other executes generation. The MCP is what removes the copy-paste step between them.
Build a persona skill and invoke it by name
The demo defines a marketing-lead agent as a skill in a repository and then invokes it in one sentence: <em>use the [named] agent to create a full marketing plan and all the creative assets I would need, using the [service] MCP server, for [this product]</em>. The agent decides which assets the brief requires and generates them.
Expect the plan, not just the images
After roughly fifteen minutes the output is a set of planning documents — brief, segments, channel plan, creative briefs, social posts, influencer plan — <strong>plus the generation prompts it wrote for each asset</strong>. Those prompts are arguably the more reusable artifact, since they are specific enough to re-run and adjust.
Why one aggregator instead of many keys
The service is described as doing for generative media what a model router does for LLMs: one login and one bill across many image and video models, with new ones added as they ship, instead of managing separate API keys and costs per provider.

Gotchas

The media service is a paid commercial product and the video is built around it; the MCP-as-creative-arm pattern works with any service exposing an MCP server.
The presenter grants the connector full tool and model access in the demo — "we're just going to give it the world". Scoping it is offered and skipped, which is worth doing differently.
A fifteen-minute planning run producing a large document set is a real token cost; the output is genuinely substantial but it is not a quick generation.
Generated marketing assets carry the usual provenance and disclosure questions; nothing in the video addresses them.

← Back to News digest · See also: Claude Cowork guide