MCP

Musevate MCP Server is a hosted service for generating videos from a prompt, a single image, or a set of reference images through a unified endpoint, routing across multiple AI video models.

Purpose The service is aimed at users and clients who want to invoke AI video generation capabilities through the MCP protocol. It converges calls across multiple video models into a single endpoint, so users do not need to integrate with each model separately.

Main capabilities

• Supports three generation modes: text-to-video (prompt only), image-to-video (provide imageUrl, the image serves as the first frame, and the prompt describes its changes), and reference-image-to-video (provide referenceImageUrls, up to thirty, referenced in the prompt in the order @Image1, @Image2).
• Supports endImageUrl to specify the end frame, but imageUrl must also be provided, and only some models provide this field.
• The mode is automatically inferred from the content sent, rather than set separately, so it will not conflict with the image content.
• Provides quote_video (free) for querying generation cost, and generate_video for actual generation and billing.
• Identical calls repeated within two minutes return the generation result started the first time, rather than regenerating; if you intentionally want to generate a second version, you can modify the request or pass in a new requestId.
• Uses a prepaid credit system with no subscription; pricing is determined before each generation, and if it fails for technical reasons, a refund is issued automatically.

How to use The service is hosted, with two access paths:

1. Direct connection: Add https://musevate.com/api/mcp as a custom connector. All methods require an account, and the first request returns 401 with a WWW-Authenticate header pointing to OAuth metadata, through which the client registers and redirects to musevate.com to log in. When selection is supported, it is recommended to choose "always required". An API key can also be used in settings as a static credential.
   • Claude (web, desktop, or mobile): Settings → Connectors → Add custom connector, paste the endpoint, and complete login in the browser.
   • ChatGPT: Settings → Connectors → Add custom connector, using the same URL.
   • Other clients that support streamable HTTP: point to this endpoint; if the client does not support OAuth, use the request header method instead.
2. Local bridge: Some clients can only start a command and communicate through a pipe. In that case, you can use musevate-mcp from the repository, which is a stdio server that forwards to the same endpoint. It is installed and built from this repository itself; an npm release named musevate-mcp is coming soon, and until then the above command is available, and it will continue to be available afterward. Keys need to be created at musevate.com/settings. The bridge cannot run the OAuth flow (which requires a browser), so it only accepts keys.

The bridge itself contains no logic: no pricing, no retry rules, and no judgment about tool behavior, to avoid having a second place that decides "what the cost is". It is only responsible for transporting bytes and correctly handling the frame format: notifications are answered silently; each response is re-serialized into exactly one line; gateway error pages, empty response bodies, or network interruptions are converted into JSON-RPC errors carrying the id the client is waiting for, rather than returning raw HTML on the protocol channel or hanging.

Manual check: the transport layer is pure JSON-RPC over POST, with no need to establish a session or keep a stream open. Without request headers, it returns 401 with a WWW-Authenticate pointing to /.well-known/oauth-protected-resource. examples/generate.mjs demonstrates the complete flow (balance, quote, generate, poll) in about one hundred lines of dependency-free code.

generate_video and quote_video accept images as https URLs, because tool calls are JSON objects cannot carry bytes. It is recommended to provide an image rather than describe it: asking a model to draw a specific logo will only produce something close, whereas providing the logo allows it to be followed accurately.

Use cases

• Need to generate videos directly in an MCP client (such as Claude, ChatGPT, or other clients that support streamable HTTP).
• Need unified invocation across multiple AI video models, and want to know the cost before deciding whether to generate.
• Need image-to-video or reference-image-to-video to maintain the consistency of specific visual elements (such as a logo).

Limitations

• All methods require an account and cannot be used anonymously.
• The local bridge cannot run OAuth and can only use API keys.
• endImageUrl is supported by only some models.
• Reference-image-to-video and imageUrl cannot be used at the same time.
• This repository contains only the local bridge, the manifest published to the registry, and usage documentation for the hosted server; the product source code is not included, the bridge is not a reimplementation of the product, and routing, pricing, profit floors, and refund rules are all server-side.
• The MIT license covers only the files in this repository (the bridge, documentation, and manifest) and does not constitute a license to the Musevate service, which is a paid product subject to its own terms.
• For pricing details, see musevate.com/pricing.