How to Use Wan AI (Wan 3.0) in Claude Code and Cursor?

Wan 3.0, Alibaba's video model, runs in Claude Code and Cursor through the Makernice MCP server: the agent gets a price from estimate_cost, you approve, and generate runs the job capped at that max_cost_usd. Wan 3.0 defaults to 720p, where a 5-second clip is $0.39 on price list v15 (November 12, 2026). Text-to-video, image-to-video, first and last frame and reference-to-video all work from the same chat.

How do I use an MCP server to generate Wan 3.0 in Claude Code or Cursor?

Wan 3.0 is generated over MCP by adding the Makernice server to your editor and signing in; you also need a Makernice account with balance. Claude Code: run claude mcp add --transport http --scope user makernice with the URL listed on the Claude Code setup page, type /mcp, pick makernice and approve in the browser. Cursor: install from the Cursor setup page or paste its block into ~/.cursor/mcp.json, then sign in beside makernice under Cursor's MCP settings.

Each connection gets a per-job limit, and you can add a daily limit as well. The get_balance tool reports both alongside your balance, which is handy before a long Wan 3.0 take that bills many seconds at once.

Wan 3.0 without an editor: Claude and ChatGPT connectors

If you would rather ask for Wan 3.0 in a chat window than in an IDE, Claude and ChatGPT can reach the same Makernice tools. Claude takes the server as a custom connector (how to add it). ChatGPT on the web takes it as a custom MCP server with OAuth sign-in, on plans or workspaces that permit those, and points at its own server address (how to add it); that version leaves out get_topup_link. Nothing else changes for Wan 3.0: the assistant gets a quote from estimate_cost, which does not reserve the amount, passes the figure you approve to generate as max_cost_usd, and fetches the finished clip with check_job.

How do I write a Wan 3.0 text-to-video request, and what settings matter?

A Wan 3.0 text-to-video request needs model, mode t2v and a prompt; the settings that change the clip are duration (2 to 30 seconds), resolution (480p, 720p, 1080p), aspect_ratio (adaptive, 16:9, 9:16, 1:1, 4:3, 3:4) and generate_audio. Describe the subject, the action and the camera in plain sentences, one shot per prompt. Here is a full request from Claude Code, priced before it runs.

You: Can you make a short clip of a fox walking through a snowy birch forest, filmed low to the ground? Show me the cost first.

Claude Code turns that into a request and calls estimate_cost:

{
  "name": "estimate_cost",
  "arguments": {
    "input": {
      "model": "wan-3.0",
      "mode": "t2v",
      "prompt": "A red fox trots through a snowy birch forest, snowflakes falling, low camera following at ground level, soft winter light",
      "duration": 5,
      "resolution": "720p",
      "aspect_ratio": "16:9"
    }
  }
}

Claude Code: Wan 3.0, 720p, 5 seconds: $0.39 at price list v15. Nothing has been charged. Generate it?

You: Yes please.

The same input then goes to generate, carrying your approved amount and a request_id:

{
  "name": "generate",
  "arguments": {
    "input": {
      "model": "wan-3.0",
      "mode": "t2v",
      "prompt": "A red fox trots through a snowy birch forest, snowflakes falling, low camera following at ground level, soft winter light",
      "duration": 5,
      "resolution": "720p",
      "aspect_ratio": "16:9"
    },
    "max_cost_usd": "0.394515",
    "request_id": "fox-birch-forest-720"
  }
}

Makernice computes the price once more when generate arrives. Within max_cost_usd and the connection's per-job limit, the job is created and charged that price; above either, the call is refused and nothing is made. If the agent resends after a network hiccup with the same request_id and input, it gets the original job back with no second charge.

The agent collects the clip with check_job; wait_s of up to 55 seconds keeps the call open until the job ends. Success puts a signed MP4 link in outputs[0].url, valid for 24 hours and renewed by calling check_job again, and console_url opens the run in your console.

How do I turn an image into a Wan 3.0 video, or animate between first and last frames?

Wan 3.0 animates a picture with mode i2v and the image in image_url as the first frame; for first and last frame, set mode flf and add the end picture in last_image_url, and Wan 3.0 fills in the motion between them. Neither mode takes an aspect ratio, because the clip follows the input image. Write the prompt about the change between the frames:

{
  "name": "estimate_cost",
  "arguments": {
    "input": {
      "model": "wan-3.0",
      "mode": "flf",
      "prompt": "The fox curls up in the snow and falls asleep as dusk turns the birches blue",
      "image_url": "https://example.com/keyframes/fox-standing.png",
      "last_image_url": "https://example.com/keyframes/fox-sleeping.png",
      "duration": 6
    }
  }
}

Input images have to be public https links to JPEG, PNG or WebP files no bigger than 10 MB, so upload local files first. Wan 3.0 accepts a wide range of shapes: shorter side from 240 px, longer side up to 8000 px, and width divided by height anywhere from 0.125 to 8.

How do I use reference images with Wan 3.0?

Wan 3.0 uses reference images in reference-to-video mode: mode r2v with 1 to 10 links in reference_images, for example a character from two angles plus a location. Name what each reference is in the prompt, and leave aspect_ratio out. Reference videos aren't supported, because a Makernice request has no field for a video file.

{
  "name": "estimate_cost",
  "arguments": {
    "input": {
      "model": "wan-3.0",
      "mode": "r2v",
      "prompt": "The fox from the references sits by the frozen lake from the references and watches the northern lights",
      "reference_images": [
        "https://example.com/refs/fox-side.jpg",
        "https://example.com/refs/fox-face.jpg",
        "https://example.com/refs/frozen-lake.jpg"
      ],
      "duration": 5,
      "resolution": "720p"
    }
  }
}

Does Wan 3.0 generate audio, and can it use an audio reference?

Wan 3.0 adds sound to every clip by default, and setting generate_audio to false gives a silent clip. A Wan 3.0 request can't take an audio file, so audio references, audio-to-video and lip-sync to your own recording are not supported on Makernice. If you plan to lay your own music or voice-over on top, ask for a silent clip:

{
  "name": "estimate_cost",
  "arguments": {
    "input": {
      "model": "wan-3.0",
      "mode": "t2v",
      "prompt": "One continuous take: the fox crosses a frozen stream, climbs a ridge and looks out over a valley at dusk",
      "duration": 25,
      "resolution": "1080p",
      "aspect_ratio": "16:9",
      "generate_audio": false
    }
  }
}

That request is also a long take: Wan 3.0 accepts any whole number of seconds from 2 to 30. Billing is per second: $0.0394 at 480p, $0.0789 at 720p and $0.1578 at 1080p on price list v15 (November 12, 2026). For a 25-second take, multiply the per-second price by 25, or let the agent ask estimate_cost for the exact amount. Full details are on the Wan 3.0 model page, and the Wan 3.0 vs MiniMax H3 comparison sets it against a similar model.

Why did my Wan 3.0 generation fail, and is the failed task charged?

Wan 3.0 requests are refused before anything is charged when a field doesn't fit the mode, a limit is hit or the balance is short; the agent gets an explanation as text and no job exists. Column one is the refusal's server-side name; column two is the opening of the returned text.

A Wan 3.0 job can still fail after it starts; check_job then shows status failed, billing released, and the held amount is back in your balance. Timeouts and input moderation rejections are released the same way, while a finished output blocked by output moderation is still charged and can go to support for review (failed generations). Wan 3.0 Prime shares the same modes, durations and resolutions, at its own per-second prices.