How to Use MiniMax H3 Max with Images and References in Claude Code?

To use MiniMax H3 Max in Claude Code or Cursor, connect the Makernice MCP server and ask for a clip from a still, a prompt or reference images: the agent gets a quote from estimate_cost, waits for your approval, then calls generate with a max_cost_usd cap. The default is 768p, and 5 seconds at that resolution is $0.34 (price list v15, November 12, 2026).

Getting the Makernice tools into Claude Code or Cursor

Open the Claude Code page, copy the claude mcp add --transport http --scope user makernice command with its server URL, and run it. Inside Claude Code, /mcp lists makernice; choose it and log in through the browser. For Cursor, the Cursor page has an install button plus the ~/.cursor/mcp.json entry, and you log in from the makernice line in Cursor's MCP settings.

Before the first clip, it helps to see the limits that apply. One call with no arguments returns them:

{
  "name": "get_balance",
  "arguments": {}
}

The reply includes available_usd, spent_today_usd and per_job_limit_usd, plus this connection's daily limit when one is set. Any job priced above the per-job limit is refused, whatever you approve in chat.

Using MiniMax H3 Max from Claude or ChatGPT chat

The same server also works outside coding agents. In Claude, add it under Customize → Connectors → Add custom connector and sign in to Makernice; on Team and Enterprise plans an owner adds the connector first in Organization settings → Connectors. The Claude connect page has the URL to paste. In ChatGPT on the web, add it from Plugins as a custom MCP server with OAuth, using the separate ChatGPT address on the ChatGPT connect page; your plan or workspace has to allow custom MCP servers. From there a MiniMax H3 Max request runs the way it does in Claude Code and Cursor: estimate_cost prices it without holding that price, generate refuses a quote above your max_cost_usd, and check_job hands back the clip.

How do I turn an image into a video with MiniMax H3 Max?

MiniMax H3 Max turns an image into a video with mode i2v: the still goes in image_url as a public https link, the motion goes in prompt, and duration and resolution work as for text-to-video. Image-to-video on MiniMax H3 Max takes no aspect_ratio, so the clip keeps the shape of the still. Here is the flow in Claude Code, priced before it runs:

You: Here's our garage photo at the public link. Animate it with MiniMax H3 Max: slow push-in, the headlight comes on. Price first.

Claude Code passes the still and the motion to estimate_cost:

{
  "name": "estimate_cost",
  "arguments": {
    "input": {
      "model": "minimax-h3-max",
      "mode": "i2v",
      "prompt": "Dust drifts through a shaft of light as the camera pushes in on the vintage motorcycle, the headlight flickers on",
      "image_url": "https://example.com/stills/garage-motorcycle.jpg",
      "duration": 5,
      "resolution": "768p"
    }
  }
}

Claude Code: $0.34 for 5 seconds at 768p with MiniMax H3 Max, price list v15. I haven't started anything. Go ahead?

You: Please do.

It then calls generate with the same input, your approved amount and a request_id:

{
  "name": "generate",
  "arguments": {
    "input": {
      "model": "minimax-h3-max",
      "mode": "i2v",
      "prompt": "Dust drifts through a shaft of light as the camera pushes in on the vintage motorcycle, the headlight flickers on",
      "image_url": "https://example.com/stills/garage-motorcycle.jpg",
      "duration": 5,
      "resolution": "768p"
    },
    "max_cost_usd": "0.34272",
    "request_id": "garage-moto-i2v-768"
  }
}

The price is worked out again when generate runs. If it stays within max_cost_usd and the per-job limit, the job is created and charged that amount, held until it finishes; if it doesn't, Makernice refuses and creates nothing. Repeating the call with the same request_id and input returns the existing job, which makes retries after a timeout safe.

For the result, the agent calls check_job with the job_id; wait_s lets it wait up to 55 seconds per call. On success, outputs[0].url holds a signed MP4 link good for 24 hours, a later check_job issues a new one, and console_url opens the job in the Makernice console.

How do I write a MiniMax H3 Max text-to-video request, and which settings matter?

A MiniMax H3 Max text-to-video request is mode t2v with a prompt of up to 4,000 characters; the settings that matter are resolution (480p, 768p, 1080p, 768p by default), duration (5 to 15 whole seconds, 5 by default) and aspect_ratio (21:9, 16:9, 4:3, 1:1, 3:4, 9:16). Resolution and duration set the price; the prompt and ratio don't. A vertical 1080p night shot, priced first:

{
  "name": "estimate_cost",
  "arguments": {
    "input": {
      "model": "minimax-h3-max",
      "mode": "t2v",
      "prompt": "Night market street in heavy rain, a rider on a vintage motorcycle weaves between stalls, neon reflections on wet asphalt, low tracking shot",
      "duration": 10,
      "resolution": "1080p",
      "aspect_ratio": "9:16"
    }
  }
}

Write the camera move, the subject's action and the lighting into the prompt, because MiniMax H3 Max has no separate camera or style fields on Makernice. Test the motion at 480p with the same prompt before paying for 1080p.

How do I use reference images with MiniMax H3 Max?

MiniMax H3 Max uses reference images in mode r2v: pass 1 to 4 public https links in reference_images and describe in the prompt what each one shows. Reference videos are not supported, since the Makernice request has no video field for MiniMax H3 Max.

MiniMax H3 Max checks every input image when the job is created, in image-to-video, first and last frame and reference-to-video alike. Each must be a public https link to a JPEG, PNG or WebP under 10 MB, with a shorter side of 256 px or more, a longer side of 5760 px or less, and a width-to-height ratio from 0.4 to 2.5. A very tall phone screenshot or a wide panorama falls outside that range and is refused before anything is held, so crop it first. Local files can't be sent over MCP.

Reference-to-video on MiniMax H3 Max accepts these ratios: adaptive, 21:9, 16:9, 4:3, 1:1, 3:4, 9:16; adaptive lets the output follow the shape of the references. Here is the agent pricing an 4-reference shot at 1080p:

{
  "name": "estimate_cost",
  "arguments": {
    "input": {
      "model": "minimax-h3-max",
      "mode": "r2v",
      "prompt": "The rider from the references wheels the motorcycle out of the garage into a rainy street at night",
      "reference_images": [
        "https://example.com/refs/rider-face.jpg",
        "https://example.com/refs/rider-jacket.jpg",
        "https://example.com/refs/motorcycle-side.jpg",
        "https://example.com/refs/garage-door.jpg"
      ],
      "duration": 8,
      "resolution": "1080p",
      "aspect_ratio": "adaptive"
    }
  }
}

On price list v15 (November 12, 2026) MiniMax H3 Max costs $0.0452 per second at 480p, $0.0685 at 768p and $0.1536 at 1080p; the 1080p tier is refined from 768p. Clips last 5 to 15 seconds, 5 by default, and the output includes audio. Specifications are on the MiniMax H3 Max model page; if you need 2K or 4K, those tiers belong to MiniMax H3.

What MiniMax H3 Max refusals mean in practice

A refused call creates no job and costs nothing; Makernice explains the reason in readable text rather than an error code. The first column is the refusal's internal name on the server and the second is the start of the message the agent sees.

A job that fails after it has started, for example because the still was rejected by moderation, shows status failed and billing released in check_job, and the held amount goes back to your balance.

Is MiniMax H3 Max open source, and can I run it locally?

MiniMax H3 Max has no official open-weights release that we could find: the MiniMaxAI page on Hugging Face lists MiniMax H3 and MiniMax H3-Image repositories but no H3 Max (checked 2026-11-12). Makernice runs MiniMax H3 Max as a hosted API only, so it can't run on your own machine through Makernice. If self-hosting matters, the MiniMax H3 guide covers the weights MiniMax did publish for the base model.