Connecting the Kive MCP connector to Claude turns product photography into a conversation: you describe the outcome, and Claude resolves your saved products, characters, and studios by name, writes the briefs, starts the renders, and polls them to completion. This article documents one real session, end to end – a three-scene mini-campaign for a fictional Scandinavian-minimal look, with every credit cost and render time taken from that run rather than from a feature page.
What the Kive connector gives Claude
The Kive MCP server exposes 28 tools to any assistant that speaks the Model Context Protocol, and Claude reads that list at connection time. The ones that matter for product shots map straight onto Kive's own concepts: saved products that keep their construction and labels intact, character models that stay the same person across generations, studios that hold a scene's lighting and framing, video presets that turn a still look into motion, and style references that lock a look across new frames.
The practical difference from Kive's composer is that nothing needs clicking. The workflow you would assemble by @-mentioning a character and picking a studio becomes a sentence: shoot this look on the saved character, at the stone crossing, then make it move. Claude plans the tool calls; the ChatGPT MCP guide walks the same tool loop from the other assistant's side, and the loop is identical here.
Where the credits come from
An MCP generation spends the same workspace credits as the app, at the same rates. Claude checks balances with a workspace listing, and the image and video generation tools accept an estimate-only flag that returns the cost without creating anything – worth using before a batch. Editing operations spend without an estimate step, so they are the ones to watch on a tight balance.
Setup in Claude's connector settings
Two minutes, three steps. In Claude, open the connector settings and add Kive – the server is mcp.kive.ai, and the connect-to-Claude docs list the exact clicks per surface. Approve the OAuth prompt with your Kive account; the connection inherits your permissions, so Claude can only reach workspaces you can reach, and revoking access in Kive revokes it for Claude too.
Then make the first request read-only: ask Claude to list your Kive workspaces. It confirms authentication, shows each workspace with its credit balance, and costs nothing. From there the library is searchable by conversation – products, characters, studios, boards – before anything generates.
The anchor workflow, run for real
The consistency problem is the one that breaks prompt-only tools: scene two never quite matches scene one. The fix is to make one image the anchor and derive everything else from it.
Step 1: one hero still
The session started with a single request to generate the hero look on a saved character model – a full-body editorial still in a bright minimal studio, wearing an oversized sand trench over a cream crewneck and ivory trousers. One model-anchored generation, 10 credits, ready in about 44 seconds. That image became the anchor for the rest of the session; it is also this article's hero image, unretouched.
Step 2: the same look, every scene
Each scene call then reused three things: the same character model, a studio resolved by name, and the hero still passed back as a style reference – with the garments described identically in every brief. Two scenes ran this way, one at blue hour on a wet street and one against a warm ochre backdrop. The trench, knit, and trousers came back identical in colour and construction in both, on the same woman, 10 credits each. Repeating the garment wording matters as much as the reference image: the style reference carries the look, the words pin the details. Training and reusing the character itself is covered in the character-model docs, and our character consistency guide covers where the approach still breaks.
Adding motion with video presets
Video runs through its own tool, and its inputs differ in one way that trips people up: studios do not apply. generate_product_video takes the character or product models, a motion preset resolved by name – Pushing in, Holding product, Revealing slowly – and a brief; if you want a still scene's atmosphere in the clip, describe it in that brief rather than passing a studio id.
Clips run 4 to 15 seconds, and the default worth keeping is 5 seconds, 9:16, premium, with audio – which the estimate tool priced at exactly 100 credits in this session. At that rate a clip costs as much as ten stills, so the honest sequencing is stills first: lock the look on images at 10 credits a frame, then spend the 100-credit calls only on looks that already survived review.
What the session cost
Every number below came back from the tool calls in this run, not from a pricing page summary.
| Call | Credits | Time to ready |
|---|---|---|
| Hero still (character model) | 10 | ~44 s |
| Scene still (model + studio + style reference) | 10 each | ~70 s each |
| Prompt-only still, no saved models | 4 | not run – estimate only |
| 5 s premium video with audio | 100 each | estimate only in this session |
| Image or video generation with estimateOnly | 0 | instant |
The three stills totalled 30 credits; the full mini-campaign shape – one hero, two scenes, two clips – prices at 230. On Kive's monthly credit rate of $0.02 that is about $4.60 of generation (kive.ai/pricing), and the free plan's 40 credits cover the three-still half of the workflow before any subscription. One caution transfers from every asynchronous pipeline: generation returns a job, not an image, and Claude polls until the render reports ready – so a stalled job several minutes old is a reissue, not a longer wait.
Run the anchor workflow once on your own product or character and the division of labour becomes obvious: you make the creative calls, Claude does the bookkeeping – resolving names, writing briefs, watching jobs, keeping the outfit description identical on frame after frame. That bookkeeping is exactly the part human sessions get sloppy about by scene three.
The wider automation picture – catalogue refreshes, seasonal variants, pipeline steps – is the same regardless of assistant, and the ChatGPT MCP guide maps those shapes. This session was the small version of all three: one look, three scenes, one short conversation.
