seedance-mcp

Why this exists

I asked Claude to make a short series of video reels. 61 clips and four rounds of revisions later it had burned through 10.3 million Seedance tokens. I found out what that meant from the bill: about $100.

The agent did nothing wrong. It reported tokens, and tokens mean nothing to a human. Dollars do.

So this server prices every request in dollars before it is sent, lets the cheap stuff through, and stops to ask you when it matters. Not in a log. Not on the invoice. Before the money is gone.

Related MCP server: dare-mcp-server

What you get

Quick start

1. Get a key. Sign up at BytePlus, open ModelArk, activate the Seedance and Seedream models under Model activation > Media (free, you pay per generation), then create a key under API Key Management.

2. Add the server. For Claude Code:

claude mcp add seedance --scope user -e ARK_API_KEY=your-key -- uvx --from git+https://github.com/atharva3vedi/seedance-mcp seedance-mcp --budget 50

For Claude Desktop, Cursor, Windsurf or any other MCP client, add this to its MCP config:

{
  "mcpServers": {
    "seedance": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/atharva3vedi/seedance-mcp", "seedance-mcp", "--budget", "50"],
      "env": { "ARK_API_KEY": "your-key" }
    }
  }
}

No uv? Run pip install git+https://github.com/atharva3vedi/seedance-mcp and use seedance-mcp as the command.

3. Ask for something.

make a 5 second clip of a cat surfing at sunset, draft quality first

Claude drafts it on Seedance 2.0 Mini at 480p (about $0.17), waits for it, and hands you the file.

The spend guard

Every generation is priced before it leaves your machine. Then:

How it asks

In a dialog when your client supports MCP elicitation (Claude Code 2.1.76 and later, among others). The server asks your client directly, your client shows you the dialog, and the answer comes from you. The model never sees a button it could press.

In chat everywhere else. The tool returns needs_your_ok with a one-time approval_id and the exact sentence to show you. Claude has to ask, and only after you say yes can it call again with that id. Each id is tied to that exact request: change the prompt, the model or the duration and it stops working. It expires after 15 minutes.

Chat approvals rely on the agent following instructions. Dialog approvals don't. Set --approval dialog to forbid the chat path entirely.

Batches

Before a shoot, Claude calls estimate_cost to tell you the total, then preapprove_spend to ask once:

Spend about $2.33 on two surf clips (2x Seedance 2.5, 5s, 720p, with audio)?
This month: $14.20 spent of your $50.00 budget. After this: $16.53.

You get back an approval_id that covers up to that many calls and that many dollars for 3 hours. A third clip, or a pricier one, needs a new yes.

It sees everything on your key

Before deciding, the server pulls your recent tasks from BytePlus, so clips made by scripts, other agents or another Claude window all count toward the budget. Images and older tasks live in a local ledger at ~/.seedance-mcp/ledger.jsonl, which every session shares.

Check your spend from a terminal any time:

seedance-mcp spend

seedance-mcp: this month
  $14.20 of $50.00  [######..............] 28%
  ask first above $0.50, hard cap none
    Seedance 2.5           11     $13.84
    Seedream 5.0 Pro        4      $0.36

Settings

Pass flags after seedance-mcp, or set environment variables. You can also keep them in ~/.seedance-mcp/.env (see .env.example), which keeps your key out of every repo.

What things cost

BytePlus list prices as of October 2026, for a clip without reference video:

Seedream 5.0 Pro is $0.045 per image up to 1.5K and $0.09 above. Flash is $0.018, Lite $0.035. Failed and moderated generations are free.

Estimates use the formula BytePlus bills with, (input video seconds + output seconds) x width x height x 24 fps / 1024 tokens plus one frame, and match real usage to the token. Prepaid token packs and discounts aren't reflected. If prices change before this table does, point SEEDANCE_PRICES_FILE at a JSON file:

{
  "video": {
    "dreamina-seedance-2-0-mini-260615": {
      "name": "Seedance 2.0 Mini", "max_input_video_seconds": 15,
      "rates": { "sd": [3.5, 2.1] }
    }
  }
}

rates are USD per million tokens, [without video input, with video input], keyed by sd (480p and 720p), 1080p and 4k.

Tools

Tips that save money

• Draft on Mini at 480p, then render the keeper on 2.5. A Mini draft costs about a seventh of a 720p render on 2.5.

• Reuse instead of re-rendering. Trim, re-time and re-caption clips in an editor. A re-shoot costs the full price again.

• Faces: Seedance 2.x rejects realistic human faces as image input. To keep one person across clips, make their first clip text to video, then pass that clip's video_url in reference_videos. The server knows that clip's length and prices it exactly.

• No watermark by default. The BytePlus API stamps one unless told not to, and this server tells it not to. Pass watermark=true if you want it.

• Let it wait. wait=true costs nothing extra and saves a dozen polling calls.

Development

git clone https://github.com/atharva3vedi/seedance-mcp && cd seedance-mcp
uv venv && uv pip install -e ".[dev]"
pytest

Tests run against a fake ModelArk and a fake human answering dialogs, so they never spend a cent. Billing formulas are pinned to token counts from real tasks.

Credits

Started life as a fork of kienbui-demo/byteplus-mcp and was rewritten from scratch for local use, with a spend guard.

Not affiliated with ByteDance or BytePlus. Seedance and Seedream are their trademarks. You pay BytePlus directly for what you generate.

MIT licensed.