A hosted MCP server for video generation

The tools an agent gets

Each tool is the MCP face of one REST operation, with the same inputs, prices and error codes:

• list_models: The model catalog: ids, inputs, aspects, durations and an example price in credits.
• get_credits: The balance behind the key and the plan, if any.
• quote: What a run would cost, without starting it. Call it before every generate.
• generate_image: Start an image run. Waits for the result by default and returns the image URLs.
• generate_video: Start a video run. Returns the job at once by default; follow it with get_job.
• get_job: One job, optionally held until it finishes.

Video models available through generate_video include Seedance 2.0, Kling 3.0, Grok Imagine, Wan 3.0, MiniMax H3, Seedance 2.5, Wan 2.7, MiniMax Hailuo 2.3, Seedance 2.0 Mini, Kling 3.0 Standard, Seedance 2.0 Fast, Kling 2.6 Pro and MiniMax Hailuo 2.3 Fast. The agent should read ids from list_models rather than guess them; each entry says which inputs the model takes (text, a start frame, an end frame, sound) and prices one example run.

How a video run works over MCP

Images come back in one call. Video takes one to several minutes, longer than most clients keep a tool call open, so generate_video charges the run and returns the job at once. The agent then calls get_job with wait_seconds, which holds for a while and returns early when the job is final, and repeats until the status is completed or failed.

Results are URLs, never inline video, so a long clip does not fill the context window. The agent can hand the URL to you, download it, or pass it on as the start frame of the next clip.

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "generate_video",
    "arguments": {
      "prompt": "a cyclist rides through an old town street at golden hour, tracking shot",
      "model": "seedance-2-5",
      "duration": 5,
      "aspect": "9:16",
      "resolution": "720p"
    }
  }
}

That call is 430 credits. A well-behaved agent sends the same arguments to quote first and tells you the number; a quote charges nothing.

Connecting a client

The server needs one header, Authorization: Bearer ks_live_YOUR_KEY, with an API key from your account page. Claude Code takes one command:

claude mcp add --transport http kinoseed https://kinoseed.app/api/mcp --header "Authorization: Bearer ks_live_YOUR_KEY"

{
  "mcpServers": {
    "kinoseed": {
      "url": "https://kinoseed.app/api/mcp",
      "headers": {
        "Authorization": "Bearer ks_live_YOUR_KEY"
      }
    }
  }
}

Windsurf, Claude Desktop (through mcp-remote, since its connector dialog expects OAuth) and other clients are set up on the MCP server page. The Claude Code guide walks through a project config and a first clip.

What a clip costs

There is no charge for the server, for list_models, get_credits, quote or get_job. Only generations spend credits. A draft on Wan 3.0 at 480p is 50 credits for 5 seconds; the same length on Seedance 2.5 at 720p is 430.

Every model's rate per second is in AI video API price per second, and the Seedance 2.5 API guide prices every length and resolution of the model agents ask for most.

Keeping an agent's spending in check

Video costs more per run than images, so the server makes the common agent mistakes cheap:

• Retries cannot double-charge. Identical generate calls close together return the same job, and an idempotency_key argument makes that explicit.
• Errors are readable. No credits or a bad model id comes back as a tool result with isError: true, so the agent can read insufficient_credits and stop instead of looping.
• Failures cost nothing. A failed run is refunded automatically, and the job says so.
• Keys are scoped and revocable. Give an agent a key for the work at hand and revoke it after. See Authentication.

The balance is the hard ceiling. For a softer one, put a budget rule in the agent's instructions; the agent guide has a system prompt block written for that.