generation whose id is the private library item being filled in. Poll get_generation_status until it is completed, then place generation.item.sourceId with the usual placement tools. Images and video clips need a paid plan (a free trial is refused with a message to upgrade), are charged in your workspace’s AI Credits, and are refused while the balance doesn’t cover them; sound effects count against your plan’s weekly AI generation allowance.
generate_image
Generate a landscape image (2048×1152) from a text prompt. Usually takes 30-90 seconds. Each image is charged in the workspace’s AI Credits, and generation is refused during a free trial or while the balance is empty. Once complete, place the item’ssourceId with add_overlay for an image overlay, or with add_layout (media: {type: "image", sourceId}) for a b-roll still.
The
sourceId of an image source to guide the generation — from create_source (kind: "image") after uploading, or an image item in list_librarysunburst (GPT Image 2.6, most capable, follows reference images most precisely), flare (GPT Image 2.6, faster), seedream-5.5 (Seedream 5.5, most accurate text and reference-image edits), seedream-5.0-pro (Seedream 5.0, good at text in images and at keeping a reference image’s layout) or seedream-5.0-flash (Seedream 5.0, faster)generation (id, type, status, prompt).
decompose_image
Split one image into a clean plate and transparent layers with Seedream (up to 16, or 29 onseedream-5.5), so each element (a card, a button, a person, a headline) can move on its own, for example in a parallax shot. The plate is the image with those elements removed and the space behind them filled in. Like the generate tools, it is not available in Claude’s connector. Unlike them, it finishes in the call: usually under a minute, up to a few. Each returned image is charged in the workspace’s AI Credits; refused during a free trial or while the balance is empty.
The image to decompose: an image source from
create_source (kind: "image") after uploading, or an image item’s sourceId from list_library or a generation. At least 512×512 pixels.Which elements to split out, e.g. “the pricing card, the Upgrade button and the headline”. Leave it out to split out every main element.
seedream-5.0-pro (cleanest layers), seedream-5.0-flash (faster) or seedream-5.5 (finer splits, up to 29 layers, with text kept on its own layers)base (the plate) and layers, each a private library item with sourceId and url. Each layer also has name, description, zIndex (1 sits just above the plate) and box, where it sits in the plate as fractions (0-1) of the plate’s width and height.
generate_video
Generate a short b-roll video clip from a text prompt. Takes several minutes. Once complete, place the item’ssourceId with add_layout (media: {type: "video", sourceId} over a time range) for b-roll, or with add_overlay for a video overlay; generation.item.durationMs is the clip’s length. When generation.status is failed, generation.error says why, for example a reference image that shows a person, or a finished clip blocked by a copyright check because a face looks like a famous person’s. The same request fails the same way, so change the prompt or reference before trying again. Each clip is charged in the workspace’s AI Credits, and generation is refused during a free trial or while the balance is empty.
The
sourceId of an image source to use as the clip’s first frame — from create_source (kind: "image") after uploading, or an image item in list_library. On the 2.5 models the clip keeps the image’s aspect ratio; the 2.0 models crop to 16:9. An image that shows a person, real or AI-generated, is refused, except one made with generate_image on seedream-5.0-pro or seedream-5.0-flash in the last 30 days and passed unchanged.seedance-2.5 (best quality, 1080p), seedance-2.5-4k (the same model in native 4K, about 1.6 times the credits and slower), seedance-2.5-flash (close to 2.5’s quality, twice as fast and cheaper, 720p), seedance-2.0-fast (cheaper still, 720p) or seedance-2.0-mini (cheapest, 720p).The clip’s length in whole seconds: 4–30 on the 2.5 models, 4–15 on the 2.0 models. A longer clip costs proportionally more credits and takes longer to generate.
generation.
generate_sound_effect
Generate a short sound effect from a text prompt, for when Tella’s curated catalog (list_library with scope: "default") has nothing suitable. Usually takes 10-30 seconds. Once complete, place the item’s sourceId with add_sound_effect (as sourceId, not presetId).
Returns a generation plus the weekly allowance (limit, remaining, resetAt).
get_generation_status
Poll this after a generate tool untilgeneration.status is completed or failed. While it is pending or running, wait a few seconds and call again. On completed, generation.item is the finished library item — pass its sourceId to add_overlay, add_sound_effect, add_layout or apply_video_edits. On failed, generation.error provides a sanitized, human-readable reason that may identify a content-safety rejection, a copyright check, a refused reference image or a timeout, or report a generic failure. Do not parse this text for machine handling; start a new generation.