Give your coding agent a video generator

The RelayDance Skill lets any skill-capable agent generate Seedance 2.0 videos with one sentence: text-to-video, image-to-video, multi-reference characters, and 480p-to-4K upscaling. Per-second billing, no queue, no KYC.

Install

Option 1: one command (Node 16+)

npx relaydance-skill

Installs to ~/.claude/skills/relaydance. Use --project for the current project only, or --dir PATH for Cursor / Codex skill directories.

Option 2: tell your agent to do it

Paste this into Claude Code (or any agent that can run commands):

Please install the RelayDance skill from
https://relaydance.com/skill/relaydance-skill.zip
into my skills directory.

Then: set your API key

export RELAYDANCE_API_KEY=sk-...        # macOS / Linux
setx RELAYDANCE_API_KEY sk-...          # Windows

Get a key from the RelayDance console (API Tokens). Restart your agent and you are done.

Use it in plain language

After install, just tell your agent what you want. It picks the model, submits the task, polls until done, and downloads the MP4. Try these:

• "Generate a 5 second video: a corgi surfing a wave at sunset, 16:9"
• "Use seedance to animate this image: ./photo.png, she turns and smiles at the camera"
• "Make a 9:16 vertical video of a neon street on a rainy night where a cat jumps onto a trash can, then upscale it to 1080p"
• "Make a 9:16 clip of @image1 dancing next to @image2" (multi-reference: repeat images, cite them in order)

A 5 second 480p clip typically renders in about 100 seconds; the agent waits and saves the file for you. Result URLs expire within 24 hours, so the skill downloads promptly.

Or drive the scripts directly

Every step is also a plain command, useful for CI pipelines and batch jobs:

# 1. submit a text-to-video task (returns a task_id)
python3 create_video.py "a small paper boat drifting on a calm pond" \
  --model doubao-seedance-2-0-mini-480p --duration 5 --ratio 16:9

# 2. wait for it and save the MP4
python3 query_task.py TASK_ID --download --output-dir ./videos

# image-to-video: local files are inlined automatically
python3 create_video.py "the sun slowly sets behind the horizon" --image ./ref.png

# upscale a previous result (cheap path to high resolution)
python3 upscale_video.py "https://...mp4" --target 1080p

Models and pricing

Tokens scale with resolution and duration (width x height x 24fps x seconds / 1024), so each resolution has its own per-million-token rate. When a request includes video input (video-to-video), the discounted "with video input" rate applies automatically. Billing is settled per second on actual output; any pre-charge difference is refunded. Full price list: relaydance.com/pricing

What is inside

Four small Python scripts (standard library only, no dependencies) plus a SKILL.md the agent reads. Everything talks to the OpenAI-compatible RelayDance API:

Prefer raw HTTP? The same API is plain REST: API documentation.

FAQ

Does the skill cost anything?

No. The skill is free and open source (MIT). You pay only for video generation through your RelayDance API key, per second of output.

Which agents are supported?

Anything that reads skill directories: Claude Code, Cursor, Codex, and compatible CLI agents. The scripts are plain Python 3, so any agent that can run a command can use them.

How long does a video take?

Typically 1 to 3 minutes for a 5 second 720p clip; 4K takes longer. The skill polls automatically.

Is this the official Seedance API?

RelayDance relays to ByteDance's official Seedance infrastructure. Output URLs are served from ByteDance TOS storage and can be verified independently.