list_models
Read model names, prices, inputs, outputs, and control limits.
READCre8vid MCP interface
Discover models, price the exact request, build the prompt, submit one controlled job, and return the result through a typed tool surface.
{
"model": "kling-3",
"mode": "image-to-video",
"prompt": "The subject turns...",
"duration_seconds": 5,
"aspect_ratio": "9:16",
"generate_audio": true,
"reference_ids": ["asset_123"],
"max_cost_usd": 0.25,
"idempotency_key": "shot-07-v2"
}Tool surface
Read first. Estimate second. Start the paid action only after the prompt, settings, and budget are explicit.
Read model names, prices, inputs, outputs, and control limits.
READPrice the exact model, duration, resolution, audio, and references.
READReturn model-specific prompt structure without starting a generation.
READSubmit the approved prompt and settings under a hard cost ceiling.
PAIDRead queue state, progress, route, cost, errors, and completion data.
READReturn final metadata and a resource link for the completed video.
READA natural-language budget is advice. A typed maximum-cost field is a limit.
MCP questions
It exposes video-generation actions as typed tools that an MCP client can discover and call. The client sends structured arguments, the server runs the video workflow, and the tool returns structured job, cost, status, or output data.
Read the full answer →Any MCP host that supports the server transport and tool calls can use the same tool schemas. Common hosts include Claude products, VS Code, Cursor, and custom applications built with an MCP SDK.
Read the full answer →Yes. Give every estimate and create tool explicit duration, output, retry, and maximum-cost fields. The server should reject a job before submission when its estimate is above the allowed amount.
Read the full answer →Use separate tools for listing models, reading one model, estimating cost, building prompt instructions, creating a job, checking status, cancelling a job, and retrieving the final result.
Read the full answer →Pass the finished prompt as a tool argument, not as hidden chat context. Include the input mode, model, duration, output, references, audio plan, final frame, and any model-specific prompt fields in the same structured request.
Read the full answer →Read available routes, controls, limits, and rates before choosing a model.
Lock the prompt, output settings, references, estimate, and cost ceiling.
Poll the job handle, then return final cost, metadata, and the video resource.
Primary sources