How to Use Kling 3.0 Pro in Claude Code and Cursor?

Kling 3.0 Pro, Kuaishou's video model, runs in Claude Code and Cursor once the Makernice MCP server is connected, with the price agreed before anything runs: the agent calls estimate_cost, you confirm, and generate uses your figure as max_cost_usd. Kling 3.0 Pro is billed at 1080p, so a 5-second clip is $0.8064 on price list v15 (November 12, 2026). The same steps work for Kling 3.0 Standard and Kling 3.0 4K with a different model id.

How do I use Kling 3.0 in Claude Code and Cursor through MCP?

Kling 3.0 reaches Claude Code and Cursor through the Makernice MCP server, a remote server you add once and sign in to in the browser. Claude Code: from the Claude Code connect page, run claude mcp add --transport http --scope user makernice with the address it shows, then in Claude Code type /mcp, choose makernice and confirm in the browser. Cursor: from the Cursor connect page, use Add to Cursor or edit ~/.cursor/mcp.json, then connect next to makernice in Cursor's MCP settings and approve.

There is no API key to paste and no Kling account to link; Kling 3.0 Pro is one of the model ids the Makernice tools accept. The connection carries a per-job limit that applies to every Kling job the agent starts, and get_balance reports it along with your balance.

How do I generate Kling 3.0 videos in Claude, ChatGPT or Codex?

Kling 3.0 videos come out of Claude, ChatGPT and Codex through the same Makernice server; only the setup step differs per app. 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.

• 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.
• Codex CLI or IDE extension: run codex mcp add makernice --url with the address on the Codex connect page, then approve the connection in the browser page that opens.
• In every app the Kling 3.0 request runs the same way: 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 write a Kling 3.0 text-to-video request?

A Kling 3.0 text-to-video request needs model, mode t2v and a prompt of up to 2500 characters; duration (3 to 15 seconds, 5 if left out) and aspect_ratio (9:16, 16:9, 1:1) are the settings that change the clip. Leave resolution out, because Kling 3.0 Pro has only one tier. An optional negative_prompt lists what to keep out of the shot. Write the subject, the action and the camera move in plain sentences. Here is the whole exchange in Claude Code.

You: Square close-up of iced lemonade on a patio for the menu page, with Kling 3.0 Pro. What's the price?

Claude Code prices the request first:

{
  "name": "estimate_cost",
  "arguments": {
    "input": {
      "model": "kling-3.0-pro",
      "mode": "t2v",
      "prompt": "Close-up of a glass of iced lemonade on a sunny patio table, condensation runs down the glass, ice cubes clink as a slice of lemon drops in, slow push-in, shallow depth of field",
      "duration": 5,
      "aspect_ratio": "1:1"
    }
  }
}

Claude Code: $0.8064 for 5 seconds of Kling 3.0 Pro at 1080p, price list v15. Nothing charged. Shall I go ahead?

You: Go ahead.

Claude Code sends generate the same input with the approved max_cost_usd and a request_id:

{
  "name": "generate",
  "arguments": {
    "input": {
      "model": "kling-3.0-pro",
      "mode": "t2v",
      "prompt": "Close-up of a glass of iced lemonade on a sunny patio table, condensation runs down the glass, ice cubes clink as a slice of lemon drops in, slow push-in, shallow depth of field",
      "duration": 5,
      "aspect_ratio": "1:1"
    },
    "max_cost_usd": "0.8064",
    "request_id": "lemonade-patio-square-1"
  }
}

generate gets a new quote at the moment it is called. The job is created, and charged that quote, only when the quote is at or below max_cost_usd and within the connection's per-job limit; otherwise the call is refused and there is nothing to pay. A repeat of the identical call with the same request_id returns the job it already created.

Then the agent waits on the job, up to 55 seconds per call:

{
  "name": "check_job",
  "arguments": {
    "job_id": "gen_...",
    "wait_s": 55
  }
}

Once status reads succeeded, outputs[0].url is a signed MP4 link valid for 24 hours; a later check_job returns a fresh link and console_url opens the job in your console. Account webhooks are delivered for MCP jobs as well.

How do I turn an image into a Kling 3.0 video, or use first and last frames?

Kling 3.0 turns an image into video with mode i2v and the picture in image_url as the first frame; for first and last frame, use mode flf and add the end picture in last_image_url. Both modes follow the shape of the input image, so the request carries no aspect_ratio. Kling 3.0 Pro and Standard offer both modes; Kling 3.0 4K is text-to-video only.

{
  "name": "estimate_cost",
  "arguments": {
    "input": {
      "model": "kling-3.0-pro",
      "mode": "i2v",
      "prompt": "The lemonade glass sweats in the heat, a bee circles the rim, the camera slowly orbits left",
      "image_url": "https://example.com/stills/lemonade-glass.jpg",
      "duration": 6
    }
  }
}

Images must be public https links to JPEG or PNG files of at most 10 MB, with the shorter side at least 300 px and width over height from 0.4 to 2.5; Kling sets no upper pixel limit. Local files won't work over MCP. In a first and last frame prompt, describe the change between the two pictures. Pricing a first and last frame version looks like this:

{
  "name": "estimate_cost",
  "arguments": {
    "input": {
      "model": "kling-3.0-pro",
      "mode": "flf",
      "prompt": "Lemonade pours into the empty glass until it is full and topped with ice",
      "image_url": "https://example.com/frames/glass-empty.png",
      "last_image_url": "https://example.com/frames/glass-full.png",
      "duration": 4
    }
  }
}

Can Kling 3.0 use reference images or copy motion from a reference video?

No, a Makernice API or MCP request for Kling 3.0 Pro, Standard or 4K takes neither reference images nor a reference video, so multi-reference shots and Motion Control on your own clip are not supported. A Makernice generation request has no field for a video file, which means no Kling model here can follow the motion of a clip you supply. For reference images on Kling, switch to Kling 4.0, which takes up to 10 in reference-to-video mode; Kling 4.0 vs Kling 3.0 covers the move. To steer motion on Kling 3.0, start from a frame in image-to-video and describe the movement in the prompt.

Resolution and prompt length

Kling 3.0 Pro has no resolution choice. Every clip is billed at 1080p, $0.1612 per second on price list v15 (November 12, 2026), and the request must not contain resolution. Changing quality means changing models: Kling 3.0 Standard is billed at 720p, and the Kling 3.0 Pro vs Standard comparison explains the trade-off.

Kling 3.0 Pro prompts are limited to 2500 characters, below the general limit on Makernice, so an agent that pastes a long shot list into one prompt will be refused. The Kling 3.0 Pro model page lists every accepted value.

Where can I use Kling 3.0: Kling's own site, an API or MCP?

Kling 3.0 is available from Kuaishou on Kling's own site, kling.ai, and on Makernice as an API model. On Makernice the same Kling 3.0 Pro, Standard and 4K ids run three ways from one prepaid balance and one price list: the REST API with an API key (API docs), the MCP server in Claude Code, Cursor, Claude, ChatGPT and Codex, and the Create page in the Makernice console, a browser form that runs the same models at the same price. Makernice has no editor or timeline.

Why did my Kling 3.0 generation fail, and is a failed job charged?

Kling 3.0 requests fail for four common reasons: a field the model doesn't take, an image outside Kling's size rules, a prompt or image refused by moderation, or a run that passes the 60 minute maximum. A refusal at request time creates nothing and charges nothing. Makernice returns refusals as readable text; the first column is the internal name of the refusal on the server, the second is the beginning of the message.

Kling 3.0 jobs that fail after starting are released rather than charged: check_job shows status failed and billing released, and the same rule covers timeouts and input moderation rejections. If a finished output is blocked by output moderation for a policy violation, the run is still charged; you can contact support for a manual review. Kling-specific causes, including Kling's own content rules, are in why Kling generations fail, and the charging rules are on failed generations.

Why is my Kling 3.0 video stuck in the queue, slow or rate limited?

A Kling 3.0 job moves from queued to generating, then to moderating while the finished output gets a safety check, and ends as succeeded, failed or canceled. Makernice doesn't publish a typical render time for Kling 3.0, so keep calling check_job with wait_s; a video job that runs past 60 minutes fails with a timeout and the hold is released. Two limits can look like slowness:

• Each account runs 2 jobs at the same time by default; a third Kling job is refused with "Too many jobs are running for this account right now." until one finishes.
• A burst of tool calls is rate limited with "Too many requests. Wait a few seconds and try again."; nothing is created or charged.
• While a job is still queued, cancel_job stops it and releases the held amount in full; once it is generating, it can no longer be canceled.