Higgsfield Katana: what it is, and how to use it

Katana is a Higgsfield skill for Claude that recreates a reference edit with your own characters, products, or clothing [official]. Instead of generating a clip from a text prompt alone, Katana preserves the reference’s movement, camera behavior, timing, and structure while replacing the subject, product, wardrobe, or visual style [official].

The workflow runs through the Higgsfield MCP connector inside Claude [official]. It also exists alongside the browser-based Higgsfield video editor — a multi-track timeline with AI generation backed by models such as Kling, Seedance, Veo, and Wan — so MCP automation and manual timeline work can complement each other [reported].

The key distinction: Katana is an editing tool, not a simple text-to-video generator. The reference video acts as a guide for motion and composition, so the quality of your input footage directly shapes the output [reported].

No API key is needed. In Claude, go to Settings → Connectors, add the Higgsfield connector with the URL https://mcp.higgsfield.ai/mcp, and sign in to your Higgsfield account [official]. An active Higgsfield subscription is required — and the most common setup failure is authorizing the wrong Higgsfield account, so confirm the signed-in account before anything else [official, reported].

Once connected, you invoke Katana from the chat (the /katana skill inside the Higgsfield connector) [reported]. The returned video lands both in the chat and in your Higgsfield Assets [official].

Every Katana edit starts from two inputs: a reference edit — a video, or a preset from the library — and up to five photos of the characters, products, or clothing that should appear in it [official]. Files go through the Higgsfield upload window in the chat, or arrive via a direct link to the media file; links to Instagram or TikTok profiles are not accepted [official].

Three editing directions are commonly used: personalized edits (your subject in a proven structure), style transfer, and motion transfer [reported]. The practical habit: identify each input’s role before you start — which file establishes the pacing, which image identifies the person, which shows the clothing — because a muddled brief produces a muddled edit [reported].

After delivery, refinement happens in the same chat: send follow-up instructions to adjust the result [official]. One focused change per iteration works best, since each regeneration is a fresh billable generation [reported].

You can also publish your own presets to the Katana library from the Katana page: a 3 to 30 second video for the motion, up to 40 images, and a prompt that mentions them with @. Presets go live immediately [official].

Katana uses your existing Higgsfield plan credits — there is no separate billing [official]. Claude shows the credit cost of each generation before confirmation, so check before you commit [official]. The number that matters is per-generation: corrections and retries that require new generations add to the cost [official].

Plan coverage matters in practice: one hands-on reviewer on the Starter plan had their first workflow fail because Katana selected a model the plan did not cover, and had to redo the edit on a cheaper model [reported]. Confirm your plan covers the model Katana picks before burning credits on iterations.

Moderation applies: some generations may be declined [official], and community presets can be taken down after reports [official]. Instagram and TikTok profile links are rejected as inputs — use direct media links or upload files [official].

What is unknown: per-generation credit costs are not published as a fixed price list — Claude discloses the cost at confirmation time, so the number depends on your plan and the models involved [unknown]. Whether Katana eventually ships as a standalone web feature rather than a Claude MCP skill has not been announced [unknown].