Seedance API in n8n and MCP: workflows and AI agents

An API key and a token balance. Tokens are prepaid (1 token = $0.05, top-ups from $20), so a workflow that loops by mistake stops at your balance instead of running up a bill. Keep the key in n8n's credential store or an environment variable, never in a node field or a shared export. KreatorFlow is an official partner of BytePlus, a ByteDance company, and the video models are ByteDance's Seedance family.

Add an HTTP Request node after your trigger. Set authentication to Generic Credential Type, then Bearer Auth, so n8n sends your key on every call. Turn on Send Body, pick JSON and paste the body from the next section.

duration is in seconds (Seedance 2.0 takes 4, 5, 6, 8, 10, 12 and 15); GET /v1/models lists every model id with its limits. An Idempotency-Key built from the row id means a step that runs twice still makes one job and one charge. The answer is an immediate HTTP 202 with data.job (id and status) and quoted_token_cost.

A vertical 720p clip with sound. $execution.resumeUrl is the resume address n8n creates for the Wait node, unique to each execution, so parallel runs never swap clips. webhook_url must be public HTTPS: a plain http URL, such as a local n8n on localhost, is ignored and the job runs without a callback, so local setups should poll instead.

Add a Wait node set to Resume On Webhook Call, HTTP Method POST, with Limit Wait Time on as a safety net. When the job ends (completed, failed, canceled or blocked), KreatorFlow posts an event of type job.<status> with the job in data.job. As with n8n's Webhook node, the posted event sits under body in the node's output. n8n does not check the X-KreatorFlow-Signature header, so treat the event as a notice and fetch the file with your own key.

Without a public HTTPS address, poll: a Wait node, a GET of the job and an IF node that loops until the status is final. GET /v1/jobs/{id} allows 120 requests per 60 seconds, far above a 15-second loop.

When the status is completed, GET the output with the same credential and the response format set to File; the Google Drive, S3 or YouTube nodes take the binary as-is. Send failed, canceled or blocked jobs down another branch with their errorCode. Tokens held for a failed job are given back.

Yes. Send the picture's public URL as first_frame_url and describe the motion. With aspect_ratio adaptive, the clip keeps the picture's shape. Images are not billed: the clip costs the same as text to video.

When a model should write the prompt or pick the format, attach n8n's MCP Client Tool to an AI Agent node and point it at the developer MCP server with Bearer Auth. Choose HTTP Streamable: the server does not offer the older Server Sent Events transport. Expose only list_models, generate_video, get_job and get_balance, so the agent can read costs, start a clip, check on it and see your balance.

https://kreatorflow.ai/v1/mcp is for clients that can send your key in a header. Claude Code adds it with one command; Cursor, VS Code and most other clients take the config in the next section. Claude's and ChatGPT's apps connect with an OAuth sign-in instead, so there use the connector at https://kreatorflow.ai/api/mcp, which spends plan credits rather than API tokens.

Replace the placeholder with your key and keep the file out of version control. The server is stateless, so every request is a plain POST; a 401 invalid_api_key means the key is wrong and a 403 account_not_active means the developer account needs attention.

10 tools: list_models, generate_image, generate_video, edit_video, lip_sync, list_voices, generate_speech, generate_sound, get_job and get_balance. Each calls the same /v1 endpoint as the REST API, so an agent's clip is quoted and billed like one from your code. generate_video takes prompt, model, duration, aspect_ratio, resolution, first_frame_url, last_frame_url, reference_image_urls, generate_audio and idempotency_key.

The agent polls get_job about every 3 seconds until the status is completed, then reads providerOutputUrl. get_job needs the jobs:read scope; get_balance and list_voices need models:read.

Errors are JSON with success: false and a code, over REST and MCP alike. Branch on the code.

The same as any API call, billed per second by model and resolution, with the exact quote returned before anything renders. A 5-second Seedance 2.0 clip at 720p is $0.55 as of 23 October 2026. Status checks and downloads are free. For agents, keep the balance small while testing and expose get_balance.