Generate videos with RunComfy

Use the connected RunComfy MCP tools; tool names may have a client-specific prefix. If authentication is needed, use the client's RunComfy sign-in flow. Keep credentials out of chat and files.

Plan the shot

1. Reuse the user's subject, action, camera motion, duration, aspect ratio, resolution, audio preference and budget. Clarify only the details needed to choose a valid request.
2. Search list_models by the requested family and capability. RunComfy includes Seedance 2.5, Wan 3.0 and LTX 2.5, as well as Kling models; current availability comes from the returned model list. Use list_model_categories when needed.
3. Call get_model for the exact returned model_id. Distinguish text-to-video, image-to-video, reference-to-video, first/last-frame and editing endpoints. Do not infer supported inputs from a family name.
4. Read the exact required input names and limits. Match duration, resolution, aspect ratio, audio and seed to the schema. A pricing tier label may differ from the encoded output dimensions.
5. Check get_balance. Show the complete request and available price estimate, including per-second pricing when applicable. Use existing user authorization and spending limits; obtain missing authorization before paid submission or publication of private input assets.

Submit once and monitor

• Use publicly accessible HTTPS URLs for model file inputs. Do not send local paths or private browser links. If the user provides a local file, use an available authorized upload route; otherwise explain the requirement.
• Call run_model once with wait_for_completion: false. Save its request ID and use get_model_request_status to monitor that same job.
• Poll with reasonable spacing. Do not resubmit just because generation is slow, the client times out, or the response is interrupted. Reconcile the existing request first.
• Fetch get_model_request_result after completion. Use cancel_model_request if the user cancels or an agreed stopping condition is reached.
• Inspect the cancellation outcome: an in-progress or finished request may be not_cancellable. Report that result and continue reconciling the job. A monitoring deadline does not guarantee cancellation or a hard spending cap, and already-incurred charges remain possible.
• Return the actual video URL and request ID. Open or embed it if the client supports video. When inline playback fails, provide the exact direct output link.
• Distinguish a completed job, successful file retrieval and actual playback verification. Report observed duration, resolution and audio only if inspected; avoid claiming visual quality from status alone.

Useful requests

• “Compare Seedance, Wan and LTX for a five-second product clip. Show current prices and supported inputs; do not run yet.”
• “Animate this mug image with a slow camera orbit. Review the exact model, duration and cost first.”

AI video models · Model API documentation