Sometimes the fastest way to prototype a video idea is not to open a separate generation dashboard, but to let your assistant create and poll the task while you stay in the conversation.
This guide shows how the Seedance MCP connector works inside Claude.ai, based on the Ace Data Cloud documentation. The workflow is practical: add one remote MCP URL, authorize with OAuth, ask Claude to list available models and resolutions, then generate a short video and poll the task until a URL is returned.
Seedance is ByteDance’s Doubao video generation family, exposed here through a remote MCP server. Once connected to Claude.ai, the documented tools include:
| Tool | Purpose |
|---|---|
seedance_generate_video |
|
| Generate a video from a text prompt | |
seedance_generate_video_from_image |
|
| Animate an image into a video | |
seedance_get_task |
|
| Query a single generation task | |
seedance_get_tasks_batch |
|
| Query multiple tasks together | |
seedance_list_actions |
|
| List supported actions | |
seedance_list_models |
|
| List available models | |
seedance_list_resolutions |
|
| List valid resolutions |
The important part is that Claude can call the list tools first, choose valid values, submit a task, and then poll with seedance_get_task
. You do not need to manually copy model names between pages.
Claude.ai supports custom MCP connectors. For Seedance, the connector uses OAuth direct connection, so the guide does not require copying an API token into Claude. You add a custom connector with this URL:
https://seedance.mcp.acedata.cloud/mcp
In Claude.ai, open settings for connectors, add a custom connector, and use a name such as:
Seedance
Then fill the URL field with:
https://seedance.mcp.acedata.cloud/mcp
After clicking connect, Claude redirects to Ace Data Cloud for OAuth authorization. Once approved, the connector appears in Claude with the Seedance tools listed above.
Video generation usually involves more than one tool call. A careful prompt may ask Claude to:
seedance_list_models
,seedance_list_resolutions
,seedance_generate_video
,seedance_get_task
.The source guide recommends setting tool approval to “Always allow” for the connector if you are comfortable with the workflow. Otherwise, Claude may ask for confirmation several times during one generation. If you are testing a new prompt or model, leaving approvals on can still be a reasonable safety choice.
A good first prompt should be specific about cost, duration, and resolution constraints without pretending you already know every valid option. The documentation uses this style:
Use Seedance to generate one short video. First list models and resolutions, pick the FASTEST/CHEAPEST model (e.g. lite/fast variant) and the LOWEST resolution. Use the SHORTEST duration (4-5 seconds). Subject: a fluffy white kitten chasing a red leaf falling in a sunny garden, side view, simple background.
In the documented run, Claude first called seedance_list_models
and seedance_list_resolutions
, then selected:
doubao-seedance-1-0-pro-fast-251015
480p
4 seconds
16:9
24fps
Then it submitted the generation with seedance_generate_video
and used seedance_get_task
to wait for the result.
That order is worth copying. Instead of hard-coding values that may become stale, let the assistant ask the MCP server for valid options first.
The same connector also supports image-to-video generation. The documented tool for that is:
seedance_generate_video_from_image
A prompt can be as direct as:
Take this image https://cdn.acedata.cloud/foo.png and animate it with Seedance fast model: slow zoom-in, soft wind, 4 seconds, 480p.
This is useful for product mockups, social clips, or turning a static storyboard frame into a short motion test. The important fields to keep explicit are the source image URL, motion direction, duration, and resolution.
When you are not sure which model variant fits a scene, use Claude as a small experiment runner. For example:
Generate the same prompt "a samurai walking through bamboo forest in mist" with Seedance pro AND fast variants at 720p, 5 seconds, and tell me which one looks better.
Claude can call seedance_list_models
, submit separate tasks, then compare the returned videos. This is especially useful during prompt exploration, where you care more about learning what works than producing the final asset in one shot.
For early ideation, batch generation can save time. The documented batch-style prompt is:
Generate 3 Seedance videos in parallel: 1) puppy on beach, 2) hot air balloon at sunrise, 3) coffee being poured. All 480p, 4s, fast model. Then return all three URLs.
Here Claude can submit multiple tasks and use:
seedance_get_tasks_batch
to check them together. I would keep batch prompts small at first: three short clips are easier to review than ten, and the feedback loop stays manageable.
A practical workflow for short-form video prototyping looks like this:
This keeps the assistant useful without turning the process into a black box. You still decide the creative direction; the MCP connector handles valid parameters, task submission, and polling.
Full setup reference: https://platform.acedata.cloud/documents/claude-mcp-seedance