Ratios, durations and required fields differ from model to model, and the awkward part of driving a generation API from an agent is that the model guesses them. Here runway_listModels returns the exact valid values for each, and every payload is checked against the OpenAPI-derived constraints before it goes out, so a wrong enum comes back as a readable local error rather than a spent request. The output links expire after 24 hours and nothing recovers them, so treat downloading as part of the job.
A client for Runway's generation API: video from an image and a prompt, images from a prompt with references, video upscaling and prompt-based editing, plus the task polling and credit balance around them.
- Video generated from an image and a text prompt, at one of the ratios the chosen model accepts
- Images generated from a prompt, optionally guided by reference images — previously generated ones or URLs, since uploads do not work as references
- Video upscaled to a higher resolution from a URL or a data URI
- Video edited from a text prompt, with reference images, through Aleph
- Spoken audio generated from text, with a voice option
- The models available per capability with the recommended default for each, and the exact valid ratio, duration and required parameters for every one of them
- Payloads validated locally against per-model constraints derived from the OpenAPI spec — allowed enums, numeric ranges, string lengths, required fields — so an invalid request fails readably without reaching the network
- Task status polled until it succeeds and carries a url, or fails and carries an error, and tasks cancelled
- The organisation's credit balance and usage details
A Runway developer account with billing set up and an API key, passed as RUNWAYML_API_SECRET. Node.js, then npm install and npm run build in the clone; the tidiest path is installing that folder as an unpacked Claude Desktop extension and pasting the key into its settings rather than editing config by hand. Generations run long enough that the documented Claude Desktop config sets MCP_TOOL_TIMEOUT explicitly. The same code runs as an HTTP server for remote clients, where REQUIRE_AUTH=true makes each user send their own key as a bearer token and a request without one gets a 401. Plan for one thing: generated images live at their returned link for 24 hours and cannot be recovered afterwards, so download them.
One command plus a key — npx -y runway-api-mcp-server, then supply credentials