ModelRunner MCP Server
The ModelRunner MCP server gives Claude, Cursor, Windsurf, and any other MCP client access to 103+ AI models — image, video, audio, 3D, and text — as callable tools through one config entry, authorized via OAuth with pay-per-use billing.
Unlike single-vendor MCP servers, one setup exposes the whole catalog: ask for a hero image, a five-second clip, a transcription, or a 3D mesh, and your assistant picks the model and returns a hosted URL in the thread.
Set up in one step
Claude Code
claude mcp add --transport http modelrunner https://mcp.modelrunner.run/mcpClaude Desktop, Cursor, Windsurf (config JSON)
{
"mcpServers": {
"modelrunner": {
"command": "npx",
"args": ["-y", "mcp-remote", "https://mcp.modelrunner.run/mcp"]
}
}
}No API key goes in the config — your browser opens once to authorize via OAuth. See the MCP setup guide for per-client details.
What your assistant can do
Discover & inspect
search, list_models, recommended_models, get_model — find the right model and read its real input schema before running it.
Run & fetch results
run_model, wait_for_request, get_request — run any model or wrapper by its owner/alias endpoint and get hosted output URLs back.
Upload files
create_upload_url / upload_file — move local images, audio, or video to a hosted URL that any model input accepts.
Author wrappers
create_wrapper, patch_wrapper, preview_wrapper — package models into reusable endpoints with prompt templates, straight from the conversation.
Example prompts: “Generate three reference frames for a misty-forest-at-dawn video” · “Transcribe this recording and summarize it” · “Turn this product photo into a 3D mesh”.
Browse the catalog by capability
Frequently asked questions
What is the ModelRunner MCP server?
It is a hosted Model Context Protocol server at https://mcp.modelrunner.run/mcp that exposes the entire ModelRunner catalog — image, video, audio, 3D, and text models — as tools any MCP client (Claude Code, Claude Desktop, Cursor, Windsurf) can call. Your assistant can search models, inspect their inputs, run them, and get hosted result URLs back in the conversation.
How is it different from a single-vendor MCP server?
Vendor MCP servers expose one company’s models. The ModelRunner MCP server exposes every model in the catalog across providers through one config entry — your assistant picks the right model per task (or you name one), and new catalog models become available without touching your setup.
Do I need an API key in my MCP config?
No. The server uses OAuth 2.1 — the first time your client connects, a browser window opens to authorize your ModelRunner account. No credentials are pasted into config files, and generation spend is governed by your account balance.
What can my assistant actually do with it?
Discover models (search, list, recommendations), read a model’s input schema, run any model or wrapper with run_model, wait for and fetch results, upload local files for URL inputs, and author reusable wrappers — the same capabilities as the REST API, expressed as tools.