Higgsfield Publishes Its Claude Ad Production Workflow

Higgsfield has published a workflow that runs advertising from planning to publication inside one Claude conversation. The two connectors are Higgsfield MCP, which attaches video and image generation models, and Meta Ads MCP, which attaches the ad account. With both hooked up, script writing, video generation, variant production, TikTok publishing and paid campaign setup all continue in the same window. Only uploading the finished video into Meta Ads Manager stays manual. The company wrote in its documentation that it verified this process as of September 17.

The process breaks into six stages. Creative research, concept and script, asset production, resizing and versioning, publishing, and the performance loop. Higgsfield explained that these six stages have traditionally been scattered one by one across different roles and tools. With the connectors attached, four of them, concept, assets, versions and performance, move entirely into the conversation, and only research and publishing stay half outside.

The company drew the boundary first. Higgsfield wrote in its official documentation, "Claude alone produces text. Scripts, hooks, briefs, and analysis of the metrics you export for it." Claude does not render video or audio itself. Higgsfield MCP has to be attached before more than 30 video models, images up to 4K resolution and voiceovers come out of the same conversation.

The model list does not belong to one company. The same connector opens Veo 3.1, Sora 2, Kling 3.0, Seedance 2.0, Wan 2.6 and MiniMax Hailuo, with Higgsfield's own Soul, Soul Cinema and Cinema Studio standing beside them. On the image side, Soul 2.0, Nano Banana Pro, Flux 2.0 and Seedream 4.5 are called through the same pipe. By default Claude picks the model that fits the shot, and you can also name a model in the prompt or send the same brief to several models at once and compare what comes back.

There is a separate device for holding a campaign's face steady. A Soul character is trained once on 20 to 80 photos and then called up as the same person in every creative that follows. Trained characters run on the Soul 2.0 and Soul Cinema models, and one generation takes one character. The company wrote that every output should be reviewed, and attached the condition that a real person's identity be trained only with their consent.

The saving shows up where variants are made. Ad Multiplier takes one approved video of 4 to 30 seconds and turns it into several independently edited versions. The two Genjutsu models go further: one replaces the product, garment or character inside a video using reference images, and the other transfers motion and camera movement from a driving clip onto new subjects. METAL has reported that Genjutsu arrived as a way to keep the performance and the camera while swapping the background. Because the approved source sets the structure, one master ad multiplies into a set with different products, people and hooks.

Publishing splits in two here. For organic, TikTok is the only network the connector posts to directly. A finished file from the Higgsfield account either goes out as a direct post or lands in TikTok drafts, with one review step for the caption, privacy settings and required declarations before that. Every other channel runs through Supercomputer on higgsfield.ai. More than 30 connectors attach to the same account, Scheduled Tasks publish to connected TikTok, LinkedIn, X and Threads accounts, and the Instagram connector is open in select regions. Finished files also land in team apps such as Slack, Google Drive and Notion.

The paid side is handled by Meta Ads MCP. Its actions come in four groups. Campaign setup creates campaigns, ad sets and ads in the ad account, always seated in a paused state. Activation is a separate confirmed step, so no budget goes out before approval. The account data group handles catalog uploads, A/B tests and reporting over the same connection. The creatives group covers single-image link creatives and boosting existing Instagram posts. The video has to be downloaded from Higgsfield assets and uploaded into Ads Manager by hand before the campaign built in the session runs on it.

Performance comes back along two paths as well. For paid campaigns, Meta MCP pulls reporting and insights into the conversation and Claude turns them into the next brief. Which hooks held attention, which variants to retire and what to generate next are written into that brief. Analytics for organic posts stay inside the connected platforms and enter the session as export files. As a filter before any spending, Virality Predictor scores a finished video on hook strength, retention risk and engagement.

The company described autonomy in three notches. Assistant writes only scripts, hooks and briefs with no connectors, and a person carries them into production tools, which suits teams that already have a production pipeline. Operator produces finished creatives through Higgsfield MCP while a person downloads and publishes them. Agent runs generation, publishing and performance retrieval in the same session, with budget limits set in the ad account and a review gate on activation, aimed at teams testing creative at volume.

The attaching itself is short. Use the connect button on higgsfield.ai/mcp, or paste the server address straight into the connectors section of Claude's settings and sign in with a Higgsfield account. An active subscription is required and there is no API key to manage separately. For paid distribution, Meta's hosted ads server is added the same way and approved through the Facebook Login for Business dialog. Custom connectors work on every Claude plan, but the free plan is limited to one, so this two-connector setup needs a paid plan. On Team and Enterprise accounts an owner has to add the connector to the organization before members can connect.

The same company's connector guide that METAL read says this server is not exclusive to Claude. Cowork, OpenClaw, Hermes Agent and NemoClaw attach to the same server, and on the Claude side web, desktop, mobile and Claude Code all come in. A team mixing several agents still works against one asset library and one credit pool. The same document wrote that production has long looked like an assembly line. One tool for writing, one for design, one for video and one for distribution stood in a row, and time and context leaked at every handoff between them.

The company listed what stays in human hands in the same document. Research is automated only as far as the Meta Ad Library, and other ad libraries are not connected. Generations through an agent deduct credits at standard rates regardless of plan, and unlimited model access applies only on higgsfield.ai. There is no built-in credit cap on the agent, so the working control is a prompt-level instruction to state the cost and wait for confirmation before generating. Higgsfield wrote, "New campaigns arrive paused, the activation step starts spending budget," and noted alongside it that Meta's documentation places responsibility for agent actions on the account holder.

METAL has reported on a case where ChatGPT drove Runway and Blender through connectors. The difference between then and now is not generation but spending. Calling a model had already moved into the conversation, and what crossed over this time is the seat where a campaign is placed in an ad account and performance numbers are pulled back. That is why every device Higgsfield planted through the document has the same shape. Campaigns are built stopped, a person presses activate, the video upload is left by hand, and permissions are set by account settings. The design centre of this workflow is not how fast the agent makes things, but where the moment it spends money is placed.