ModelRunner

What the ModelRunner MCP server does

The ModelRunner MCP server exposes ModelRunner’s hosted catalog through the Model Context Protocol. The catalog contains more than 100 models for text-to-image, image-to-image, text-to-video, image-to-video, video-to-video, music generation, speech-to-text, and image-to-3D tasks. Models from multiple AI providers are available through one endpoint and one pay-per-use billing account.

An MCP client can discover models, inspect their input requirements, submit inference requests, and retrieve results as hosted URLs. The service also supports reusable wrappers: simplified interfaces built over base models with selected inputs, prompt templates, and predefined settings.

How it works

The remote endpoint is https://mcp.modelrunner.run/mcp and uses Streamable HTTP. Authentication uses OAuth 2.1. During first use, the client opens a browser for ModelRunner sign-in and authorization; the setup does not require copying an API key into the MCP configuration.

Model discovery can use search, model and wrapper listings, recommendations, and detailed metadata. Schema-related tools expose raw model or wrapper input definitions, prompt templates, and field mappings. After a run starts, the client can wait for completion, inspect a request, or list the user’s previous requests.

File-based workflows can request an upload URL and upload a file before running a model. Wrapper authors can consult the authoring guide, create or update wrappers, preview them, list their own wrappers, and delete them. Runs use prepaid credits and are charged per use rather than through a subscription; exact pricing varies by model and is shown on its model and pricing pages.

Setup and configuration

The ModelRunner MCP server is hosted by ModelRunner, so there is no local package to install and no server process to self-host. Add the remote URL directly in clients that accept remote MCP URLs. For clients that only support stdio, use mcp-remote as a bridge to the same endpoint.

Claude Code can be configured with its HTTP transport command. Cursor and VS Code can use a server entry containing the URL. Claude Desktop and other stdio-only clients can launch npx -y mcp-remote with the endpoint as its argument. The first connection requires browser authorization.

Tools and capabilities

The ModelRunner MCP server groups its tools into four areas:

• Discovery and inspection: search, list models and wrappers, get recommendations, inspect individual records, read raw schemas, and retrieve wrapper templates or field mappings.
• Execution and retrieval: run a model, wait for a request, inspect a request, and list the user’s requests.
• File handling: create an upload URL and upload a file.
• Wrapper authoring: read the authoring guide, create, patch, preview, list, and delete wrappers.

These capabilities cover both direct model execution and wrapper-based execution through the same run-oriented workflow.

Limitations and notes

Access is free to connect, but model execution consumes prepaid credits. Pricing is not uniform: a model may charge by image, video duration, megapixel count, or another model-specific unit. Review the individual model page or pricing page before running paid workloads.

The service is remote and first-party hosted. The repository documents the hosted endpoint rather than providing a self-hostable server implementation. Available model types and wrapper behavior depend on the catalog and the schemas exposed by each entry.