video_generate
Generate short video clips from text prompts, optionally guided by start/end frames or reference media, with batching for multiple clips and job status tracking.
Instructions
For a user's video request, load the generating-videos skill BEFORE calling this — it picks the right model and builds the motion prompt (this tool does neither, and calling it raw gives weak, generic clips). Generate one or many short video clips from text prompts, optionally guided by a start (and end) frame or by reference images, videos, or audio. Pass requests: ONE object per clip (wrap even a single clip — { requests: [ { prompt } ] }); add more objects (up to 10) to batch DIFFERENT clips in one call, and repeat an object for variations of one prompt — a single approval covers the batch. Models differ in the aspect ratios, durations, resolutions, and media they accept — call list_video_models to check. Video generation is long-running: each clip is submitted and polled for you. A clip that finishes in time returns a hosted video URL plus a local file path; a clip still generating returns {status:"pending", ...} (a job handle, NOT an error) — pass that exact handle to job_status to retrieve it, and never re-submit a pending clip. Set dry_run=true to preview the exact requests without generating (no credits spent).Each entry in results is one of three things: finished media; a {status:"pending", ...} job handle to rejoin with job_status; or a failure carrying ok: false and an error. A failed entry is terminal — report its error and never poll or re-submit it. Read every entry rather than the top-level counters alone.
Input Schema
| Name | Required | Description | Default |
|---|---|---|---|
| dry_run | No | If true, return the requests that would be sent (keys masked), make no API call. | |
| requests | Yes | One object per clip (wrap even a single clip); add more objects to batch different clips in one call (up to 10). |