MCP

MCP server

Use MLMux from AI assistants and coding agents such as Claude Code, Cursor, and VS Code through the Model Context Protocol: list models, estimate costs, generate images, and create videos.

Endpoint and authentication

The MLMux MCP server is a remote server that speaks the Streamable HTTP transport. Add it to your client with this URL:

Streamable HTTP

https://www.mlmux.com/api/mcp

Authenticate with your MLMux API key in the Authorization: Bearer header; create one under API keys. The server passes the key to the MLMux API for each call and does not store it. Without a key, only list_models and estimate_cost work.

Connect a client

Pick your client and add the server. Each example reads the key from your environment or asks for it, so it never sits in a config file you might share.

claude mcp add --transport http mlmux https://www.mlmux.com/api/mcp \
  --header "Authorization: Bearer $MLMUX_API_KEY"

Set MLMUX_API_KEY in your shell first, then run the command once.

Any client that supports remote MCP servers over Streamable HTTP with a custom header works the same way.

Tools

The server offers seven tools. Clients show them to the assistant, which picks one based on what you ask.

Tool	API key	Description
`list_models`	Not needed	Lists image and video models with their model IDs, tasks, sizes, and estimated prices on each route.
`estimate_cost`	Not needed	Estimates the cost of a number of images or video clips on both routes, with the same prices as the price calculator.
`generate_image`	Required	Generates images from a prompt and returns their URLs. Charges your balance.
`edit_image`	Required	Edits or combines reference images (public URLs) with a prompt and returns the new image URLs. Charges your balance.
`create_video`	Required	Starts a Seedance video task (text to video, image to video, or first and last frame) and returns its task ID. Charges your balance when the video succeeds.
`get_task`	Required	Returns a task's status and, once it has succeeded, the output URLs and the amount charged.
`get_balance`	Required	Returns your credit balance.

Billing

generate_image, edit_image, and create_video are billed exactly like the same API requests, at the rates on the pricing page: 1 credit = 1 US dollar, and failed requests are not charged. Most clients ask you to approve a tool call before it runs; ask the assistant to run estimate_cost first when the cost matters.

Limits

Image tools wait up to about four minutes for the result. A slower image returns a task ID; ask for get_task to fetch it later.
create_video returns right away with a task ID. The assistant then calls get_task until the video is ready.
Requests through MCP count toward your account's limit of 60 tasks running at the same time.
Sign-in with OAuth is not available yet, so connectors that require it, such as those in the claude.ai and ChatGPT web apps, cannot connect for now. Clients that accept a custom header work today.