How to Use Higgsfield Katana in Claude (2026 Guide)

Higgsfield released Katana on October 8, 2026. It is a skill for Claude: you give it a reference edit and up to five photos of your people, products or clothes, and it recreates that edit with your inputs, following the reference's shot order, pacing and cuts. The video is made inside a Claude chat through Higgsfield's connector, not on Higgsfield's website, so setup starts in Claude's settings.

Every fact about Higgsfield here comes from their own Katana page, their Katana blog post (October 12), their changelog and their official X account, read on October 13, 2026. We have not run Katana ourselves, and where Higgsfield's pages are silent or disagree, this guide says so instead of guessing.

A disclosure before we start: we make Riffkit, which has its own reshoot mode, Swap, and it also works inside Claude. If what you want is a video of your own that already performed, back with a different person, product, outfit or setting, Swap reshoots it at its full length (currently up to 45 seconds) and changes only what you name. Every new account gets free credits with no card: a short clip comes back whole, and most longer ones start with a free first section that a plan finishes from the same script. If the winner is someone else's, its Adapt mode keeps the beats and rebuilds it as a new video, which is the safer route on TikTok Shop.

What Katana actually does

Higgsfield's blog puts it in one line: "Katana is how Higgsfield handles editing inside Claude." The Katana page says: "Pick a reference edit. Katana recreates it with your inputs, from the pacing and transitions to the visual style." The launch post on X adds the kinds of edits it is aimed at: "editable motion graphics, product launch videos, or aura-farming edits."

The word to hold on to is edit. Katana is pitched at a finished, styled edit with transitions and graphics, not a single raw shot. The blog notes that "Generating the clips is only part of the work": Katana also follows the reference's shot order, pacing and cuts.

What comes back, per the blog: an MP4 with the duration, aspect ratio and frame rate of the reference, and "these parameters are not set separately in Claude." The result appears in a widget in the chat with a preview and a download button, and it is also saved to your Higgsfield Assets.

Where it runs

Katana runs in Claude through the Higgsfield connector. Higgsfield's blog describes claude.ai and Claude Desktop. There is also a web form at higgsfield.ai/katana, but it does not generate anything itself: it collects your reference, your photos and an optional note, then hands the brief to a Claude chat ("Send & Generate in Claude"). It uses the same Higgsfield account and credits either way.

Connecting it, from Higgsfield's blog:

1. Open claude.ai or Claude Desktop and go to Settings, then Connectors.
2. Click Add custom connector.
3. Name it Higgsfield and paste the server URL https://mcp.higgsfield.ai/mcp.
4. Click Connect. A Higgsfield authorization window opens.
5. Sign in to your Higgsfield account and authorize. The connector then shows in your active connectors.

No API key is involved. If you want Higgsfield in Claude Code, Cursor or ChatGPT as well, we covered each of those setups in how to use Higgsfield in Claude, Cursor and ChatGPT. Higgsfield's Katana post itself only describes claude.ai and Claude Desktop.

Three ways to start an edit

Once the connector is on, the blog lists three ways in:

• /katana opens the preset catalog in the chat.
• /katana/<preset-name> runs a specific preset.
• A plain request, such as "make me an edit of this product using Higgsfield", starts it with no command at all.

Presets come in three categories. Personalized edits "place your characters or products into a finished edit", and the other two are Style transfer and Motion transfer. Presets come from Higgsfield and from the community, and the blog notes that counts may change. Higgsfield's changelog names the styles differently, as Aura Farming, Motion and Clipping, so expect the labels to move while the product settles.

You can publish your own community preset from the Katana page. Per the blog, a shared preset needs "a 3 to 30 second video for the motion, up to 40 images, and a prompt that mentions them with @." Shared presets "go live right away and can be taken down after reports."

What you need before you start

• An active Higgsfield subscription. The blog: "It requires an active Higgsfield subscription and no API key."
• Claude, with the connector added (claude.ai or Claude Desktop, as above).
• A reference edit: a preset from the catalog, or a video of your own. Higgsfield asks you to "use a reference edit that you own or have permission to use", "including any music, footage, characters, logos, or other protected material in it."
• Up to five photos "of the people, products, or clothing that should appear in the edit." The Katana page asks for JPEG, PNG or WebP images.
• An upload, not a social link. "Links to Instagram or TikTok profiles are not accepted; the reference has to be uploaded as a file." A direct link to a media file also works.

One detail that saves a wasted turn: images you attach directly in the Claude chat "are not passed to Higgsfield MCP as Katana reference files." Katana opens a Higgsfield upload window in the conversation instead, and that is the route for your photos.

Your first Katana edit, step by step

The changelog's short version is: open the Katana page, pick a preset, upload your images, then Recreate in Claude. Here are the same steps with what matters at each one.

1. Pick the reference. A preset if you want a known style, your own video if you already have the edit you want remade. On the web form this is "Add a reference video".

2. Add your photos. Up to five, in the Higgsfield upload window, not as chat attachments. The web form labels this field "Characters, products or clothes". Clean, single-subject photos make the job easier for any generator: one person facing the camera, one product on a plain background.

3. Say what goes where, if it is not obvious. The instructions field is optional and capped at 2,000 characters. If you upload a person and a product, say which one replaces what. Name the change rather than describing the whole video again; we explain why in write what changes, not a prompt.

4. Read the price card. "Before each generation, Claude shows a card with the exact credit cost, so the price is visible before confirmation."

5. Review and correct in plain words. "Corrections are sent as plain messages in Claude: ask to change a shot, swap a photo, or adjust a detail." In Higgsfield's own test, "only the affected shots were regenerated." A correction that needs new generations can cost more credits.

6. Make more versions of what works. The Katana page invites you to "run many versions from one setup, then generate more of whatever works."

What it costs

Katana has no price of its own. "Katana uses the credits of your existing Higgsfield plan," and "there is no separate billing for MCP and no markup for running Katana inside Claude." Claude shows the exact credit cost before each generation.

Three things decide what an edit really costs you:

• Corrections. Asking for a shot again is a new generation, and "additional credits may apply."
• Extra versions. Each version you try is a new generation, and Higgsfield's blog says additional generations or alternatives may increase the cost. Budget for more than one render.
• Moderation. "Some generations may be declined by moderation." Higgsfield's Katana pages do not say whether a declined generation is charged.

Higgsfield publishes no per-edit credit figure for Katana, and its blog says plan tiers and credit allowances differ by region. Check higgsfield.ai/pricing yourself rather than trusting a number quoted in a blog post, including this one.

What the documentation does not answer

We looked for each of these on Higgsfield's own pages and did not find a clear answer:

• What does the work. The Katana page and the X launch post say "Powered by Claude Motion". The blog says Katana runs on Higgsfield models. Which parts are generated video and which are built as motion graphics is not spelled out.
• Which Claude plan you need. Higgsfield requires its own subscription and names no Claude plan.
• How long your own reference can be. The 3 to 30 second rule is written for community presets. For your own upload, no limit is stated.
• What the reference is. The changelog describes edits "from one reference photo"; the blog and the Katana page describe a reference video or preset plus your photos.
• What happens to voices. The blog says Katana "reads its shots, pacing, and audio", and in Higgsfield's own test a reference clip came back with "the same shot order, the same timing, and the same track". It does not say what happens to voices or spoken lines in the reference.
• Chat attachments. The blog says images attached directly in the chat are not passed as Katana reference files, and also that "Files attached together with the command are used as input". The Higgsfield upload window is the route both statements support.
• How long a render takes. No generation time is given.

Where a reshoot drifts

We build the same kind of reshoot into Riffkit, so this section is what breaks in our own renders, not a measurement of Katana. We have not tested Katana for any of it, but these are the things to check in any tool that rebuilds a video around new inputs.

Small details move. In our background test, the cuts landed where the original has them and the beach held across all four shots, but a necklace vanished in one shot, a sticker appeared in another and the neckline of a top changed.

Reference photos bring their whole frame. When we swapped in a saved character, her face held in every shot, and on the fire escape she also wore the jacket from her character image. Neutral photos beat styled ones.

Colour and the words. In our product swap, the new lipstick was in every beat, but its shade drifted from coral to dusty rose at about 9 seconds, and the creator still says "este glos" (this gloss) over the new product, because Swap keeps the source's words unless you ask for a change to them. Pick a source whose lines fit your product.

The habit that catches all of it: scrub every shot of the result before you post, at full size, not in the chat preview.

Before you post

Rights, and the platform rule that outranks them. Katana's own condition is a reference you own or have permission to use, music included. If you sell on TikTok Shop, the stricter rule is TikTok's: its unoriginal-content policy names "Recreating a video that directly mirrors the actions of another creator's content", and also "Using AI-generated content that results in duplicate or highly similar videos to other creators". New frames do not exempt you. Reshoot your own videos, and rebuild someone else's winner as a new video instead of mirroring it; can you legally remake a viral TikTok draws the line.

The AI label. TikTok's disclosure toggle is off by default and cannot be added after posting. Here is where it lives.

If the edit you want back is a video that already worked

Katana starts from a preset or a reference you upload and rebuilds the whole edit around your photos. Swap, the Riffkit mode from the disclosure above, is built for one job: a video that already performed, given back with your change on the source's own clock, so the shots, cuts, timing, sound and frame shape stay. The source can be a template, an upload of up to 100 MB, or a TikTok link to one video, up to 45 seconds. Pick a character, a product with photos or a reference image (an outfit or a place), or write one line in "What to change", and each engine's per-second rate shows before anything renders. It works in Claude too, through Riffkit's own connector, with no API key: replace the product in a video and replace the person in a video show the requests, and making videos in ChatGPT and Claude covers the setup.

Selling a product? Save it once with a clean photo (several on a plan) and it goes on camera in the remade video, in Swap or in Adapt. See how product placement works, or the TikTok Shop version. When you want whole videos rather than a first section, the plans are here.

For Higgsfield's tool that rebuilds part of an existing clip, see how to use Higgsfield Genjutsu. For the other ways to do this job, read Higgsfield Katana alternative.