How to Use Wan 3.0 Prime in Claude Code and Cursor?

Wan 3.0 Prime, Alibaba's video model, runs in Claude Code or Cursor once the Makernice MCP server is added: the agent requests a price from estimate_cost, you approve the amount, and generate starts the clip with that amount as max_cost_usd. A 5-second clip at the default 720p is $0.62 on price list v15 (November 12, 2026).

What does Wan 3.0 Prime need before the first job in Claude Code or Cursor?

Wan 3.0 Prime uses the same Makernice MCP connection as every other model, and the command and server address are on the setup pages. The Claude Code setup page gives claude mcp add --transport http --scope user makernice followed by the address; run it, start Claude Code, enter /mcp and approve makernice in the browser. The Cursor setup page has an Add to Cursor link and the ~/.cursor/mcp.json text; then sign in next to makernice in Cursor's MCP settings.

Signing in creates a connection with its own per-job limit and, if you choose, a daily limit. Neither can be exceeded from chat, whatever the agent proposes.

Does the Wan 3.0 Prime flow change in Claude or ChatGPT?

No, a Wan 3.0 Prime job makes the same tool calls in Claude, ChatGPT and the editors; only the setup step differs: an estimate_cost quote that informs you but locks nothing, a generate call capped by max_cost_usd, then check_job until the clip is ready.

• In Claude, open Customize, then Connectors, and choose Add custom connector; paste the MCP URL and connect with your Makernice login. Team and Enterprise workspaces need an owner to add it first. Full steps: Claude connect page.
• In ChatGPT on the web, open Plugins and add a custom MCP server with OAuth, pasting the ChatGPT-specific address; your plan or workspace must allow this. Mention Makernice with @ in a new chat to use it. Full steps: ChatGPT connect page.

How do I combine cast and set photos as Wan 3.0 Prime reference images?

Wan 3.0 Prime uses reference images in reference-to-video mode: mode r2v with up to 10 links in reference_images, enough to give it both a person and a place, and no aspect_ratio. The photos below are two of a barista and two of the cafe, all at public links.

You: Use the barista and cafe photos from the shoot folder we uploaded and make a shot of her pouring latte art at the counter. Cost first, please.

Claude Code sends the references and the scene to estimate_cost:

{
  "name": "estimate_cost",
  "arguments": {
    "input": {
      "model": "wan-3.0-prime",
      "mode": "r2v",
      "prompt": "The barista from the references pours latte art at the counter of the cafe from the references, morning rush, warm window light, medium close-up",
      "reference_images": [
        "https://example.com/cast/barista-portrait.jpg",
        "https://example.com/cast/barista-apron.jpg",
        "https://example.com/sets/cafe-counter.jpg",
        "https://example.com/sets/cafe-window.jpg"
      ],
      "duration": 5,
      "resolution": "720p"
    }
  }
}

Claude Code: Wan 3.0 Prime, 720p, 5 seconds, four references: $0.62 (price list v15). Nothing is charged until I start it. OK?

You: OK, run it.

Then the identical input goes to generate, with the approved amount and a request_id naming the shot:

{
  "name": "generate",
  "arguments": {
    "input": {
      "model": "wan-3.0-prime",
      "mode": "r2v",
      "prompt": "The barista from the references pours latte art at the counter of the cafe from the references, morning rush, warm window light, medium close-up",
      "reference_images": [
        "https://example.com/cast/barista-portrait.jpg",
        "https://example.com/cast/barista-apron.jpg",
        "https://example.com/sets/cafe-counter.jpg",
        "https://example.com/sets/cafe-window.jpg"
      ],
      "duration": 5,
      "resolution": "720p"
    },
    "max_cost_usd": "0.617145",
    "request_id": "cafe-barista-r2v-shot1"
  }
}

At that point Makernice prices the job again. A price no higher than max_cost_usd, and within the connection's per-job limit, creates the job and is the amount charged; anything higher is refused and no job is created. Calling again with the same request_id and the same input returns the original job, so retries never double-bill.

Changed your mind before it starts? While the job is still queued, cancel_job stops it and releases the held amount in full:

{
  "name": "cancel_job",
  "arguments": {
    "job_id": "gen_..."
  }
}

Once generation has begun, cancellation is refused and the job runs to the end. To get the result, the agent calls check_job, with wait_s up to 55 seconds to wait in one call. The finished MP4 is at outputs[0].url, a signed link that lasts 24 hours; another check_job gives a fresh link and console_url opens the job in the console.

How do I turn an image into a video with Wan 3.0 Prime?

Wan 3.0 Prime turns an image into video with mode i2v and the picture in image_url as the first frame; to end on a second picture, use mode flf and add last_image_url. Both modes follow the input image, so the request leaves aspect_ratio out, while resolution and duration work as in every other mode. Pricing an 8-second clip from one counter still looks like this:

{
  "name": "estimate_cost",
  "arguments": {
    "input": {
      "model": "wan-3.0-prime",
      "mode": "i2v",
      "prompt": "Steam rises from the cup on the counter as the barista slides it toward the camera, morning light flickers through the window",
      "image_url": "https://example.com/stills/cafe-counter-cup.jpg",
      "duration": 8,
      "resolution": "720p"
    }
  }
}

Which settings matter in a Wan 3.0 Prime text-to-video request?

A Wan 3.0 Prime text-to-video request needs model, mode t2v and a prompt; the settings that matter are resolution (480p, 720p, 1080p, 720p by default), duration (2 to 30 seconds), aspect_ratio (adaptive, 16:9, 9:16, 1:1, 4:3, 3:4) and generate_audio, which is on unless set to false. Describe one continuous shot per prompt, with the subject, the action and the camera.

Wan 3.0 Prime and Wan 3.0 accept exactly the same requests: 480p, 720p, 1080p, 2 to 30 seconds, the same four modes and the same image rules. Each has its own rows on the price list: Wan 3.0 Prime is $0.0617, $0.1234 and $0.2468 per second at 480p, 720p and 1080p; Wan 3.0 is $0.0394, $0.0789 and $0.1578 (price list v15, November 12, 2026). To decide on a specific shot, have the agent run estimate_cost with both model ids.

Switching off sound and raising length or resolution are single field changes. The request below turns the reference shot into 12 seconds at 1080p without sound:

{
  "name": "estimate_cost",
  "arguments": {
    "input": {
      "model": "wan-3.0-prime",
      "mode": "r2v",
      "prompt": "The barista from the references hands a cup across the counter and smiles, the cafe from the references in the background",
      "reference_images": [
        "https://example.com/cast/barista-portrait.jpg",
        "https://example.com/sets/cafe-counter.jpg"
      ],
      "duration": 12,
      "resolution": "1080p",
      "generate_audio": false
    }
  }
}

Every image must be a public https link to a JPEG, PNG or WebP of up to 10 MB; a path on your machine isn't reachable over MCP. Wan 3.0 Prime takes a shorter side from 240 px, a longer side up to 8000 px and a width-to-height ratio between 0.125 and 8. The Wan 3.0 Prime model page has the complete specification, and the reference-to-video guide covers choosing references.

Wan 3.0 Prime refusals and the fix for each

Refusals reach the agent as readable text, and nothing has been created or charged. The first column holds the refusal's internal server name; the second column holds the first words of the returned text.