Let Claude Code Make Your Videos - Kadraris MCP With a Price Check First

With the Kadraris MCP server connected, Claude Code can order a video, wait for it and drop the file into your repo next to the code that uses it. Two models are involved and they do different jobs: the one you chat with in Claude Code plans the work, and a media model such as Seedance 2.5 renders the clip. Connecting MCP does not swap out your chat model.

What you need before you start

• A Kadraris account with some balance. Connecting the server is free; only a confirmed generation costs money.
• A separate API key for the agent. Create it under API keys and give it a monthly cap.
• A place in the project for output, for example public/generated/.

Copy the key when the dashboard shows it, and keep it out of the repository and out of the prompt you give the agent.

Register the server in Claude Code

claude mcp add --transport http kadraris \
  https://kadraris.com/mcp \
  --header "Authorization: Bearer sk-kd-..."
claude mcp list

Put your own key in place of sk-kd-.... claude mcp list should now show kadraris.

Using Codex or Cursor too? They read the key from KADRARIS_API_KEY in the environment that starts the app: $env:KADRARIS_API_KEY="sk-kd-..." in PowerShell, export KADRARIS_API_KEY="sk-kd-..." on macOS and Linux. Restart the editor after you change it. For Cursor, the project config lives in .cursor/mcp.json. Every install option is listed in the MCP docs.

A prompt that quotes before it spends

The safest pattern is to make the agent show you the price and stop. A prompt I use:

Make three vertical videos for a coffee shop.
First look up the models in the kadraris MCP and
their required inputs. Get a quote for each clip
with estimate_cost, show me the total, and wait.
Generate only after I approve, using confirm_token.
Save results to public/generated/, check the file
type and size, never overwrite existing files and
never publish anything on your own.

For a first run, Seedance 2.5 is a sensible pick. Its current unit rate is from $0.261 per second; that is a per-unit price, not the cost of your whole job. Work out the full amount for your settings and number of clips in the calculator or on pricing.

The tools the agent will call

1. search_models and get_model to find a model and read which inputs it accepts.
2. estimate_cost for every planned clip, returning a quote and a confirm_token.
3. generate with the matching confirm_token, only after you say yes.
4. wait_for_task or get_task to follow progress.

Change the prompt, the model or the resolution and the old quote no longer applies: the agent needs a new quote and a new approval. Confirmation tokens also expire after a while.

After the task finishes

• Do nothing with a task until it reports succeeded. A queued task has no output yet.
• Download each URL from output into the project. The links are temporary, so never paste them into your site.
• Open every saved file, check its dimensions and watch it in full before you use it.
• Note the model, prompt, settings and cost next to the file. A clip that rendered fine but that you decide not to use is still charged; a failed generation costs nothing.

Fixing a failed connection

• Cannot connect: check the /mcp address, whether the key is active, and whether the editor process actually sees the environment variable.
• Authorization or balance error: look at the key's cap and the account balance.
• Input rejected: call get_model again for the model you are using instead of reusing settings from a different one.
• Reply cut off mid-job: look up the task id the agent saved before you start another generation, so you do not pay twice.

More agent setups are on AI agents. To give a coding agent standing instructions for the API, add AGENTS.md to the project. For pay-per-token Claude Code itself, see Claude Code with a custom base URL.

Questions

Will Claude Code generate anything without asking me?

Not if you tell it to wait, as in the prompt above. estimate_cost only returns a price and a token; nothing renders until generate is called with that token. A monthly cap on the key is a second safety net.

Is the MCP balance separate from Kadraris Studio?

No. MCP, the API and Studio all use the same account and balance. Whatever you pay for the chat model in your editor is billed separately by that service.

Why do I need a new quote after changing the resolution?

The quote covers one exact combination of model, prompt and settings. Any change alters the price, so the agent has to call estimate_cost again and you approve the new amount.