5 named agent jobs that will not start without a still in 2026

Five /agent capabilities refuse a prompt-only turn: turn a photo into a video, first-and-last-frame (two stills), transfer a dance onto my photo (still plus driving clip), run a one-tap template, and edit a photo — attach the hosted URL before the tool fires.

The in-app agent will take a sentence and start some jobs. These five will not. Each one maps to a real tool that needs a Versely-hosted image URL (and, for two of them, a second file) before dispatch. Typing “animate my product” with nothing attached is a follow-up, a failed run, or a different capability that invents the plate. Attach first. Then name the job. This list is the photo-in slice of src/lib/agentCapabilities.ts. Text-to-video and text-to-image still start from words. These five do not.

Prompt-only is a different catalog

Generate-a-video-from-text is generate_videos: prompt and models, no image required. Generate-an-image-from-text is generate_images: same shape. Those jobs ask about duration or aspect, then spend. They are allowed to start empty-handed.

The five below share a different contract. The agent gathers the still as a hosted URL — usually the chat + button — before the tool fires. A sentence that describes a photo is not an attachment. request_asset_upload is for workflow assets, not for these five.

The planner (treat /chat-models as live; the documented default is runpod/kimi-k2.7-code) can talk about the job. It cannot invent the jpeg. GLM 5.3 (z-ai/glm-5.3) is text-only: do not attach a still to that brain. Grok 4.6 (x-ai/grok-4.6) can see an attachment as a brief. Neither is a generate row. The still still has to be on the wire.

The five

1. Turn a photo into a video

Turn a photo into a video is generate_video_from_image with one still as image_url. Attach the photo, describe the motion, pick duration and an image-to-video model. Output is a clip that opens on your file. Cost is priced per model, shown before you confirm.

A prompt with no jpeg is not this capability. “Make a product orbit of my bottle” without the bottle is text-to-video guessing a bottle.

2. First-and-last-frame (two stills)

Create a first-and-last-frame transition video is the same primary tool in a different mode: first_frame_url and last_frame_url instead of a single image_url. Two photos in. One interpolated clip out. First-last frame is the mechanic: ordinary image-to-video knows where it begins and improvises the rest; first-last has to land.

Missing the last frame is a different job. One still with “and then it is empty” is an open-ended wiggle. Two stills that do not share geometry are a morph. Attach both hosted URLs before the tool fires.

3. Transfer a dance onto my photo (still plus driving clip)

Transfer a dance or motion onto my photo needs two inputs: the character still as image_url and a reference clip as video_url. The agent routes to a Kling Motion Control model. Output is your subject performing the driver’s motion. Premium-tier, priced per model, shown before you confirm.

A paragraph about choreography is not a driver. Image-to-video with extra adjectives is a pan, not a copied routine. Count the attachments: still plus clip, or it is not this job.

4. Run a one-tap template

Run a one-tap AI video template is start_template_run after every required input is a hosted URL. get_template lists what that template needs — usually a photo attached with the + button. A live card tracks the run. Each template lists its own estimated credit cost before you run it. The templates catalog is the shelf.

Launching with an empty inputs object is a failed run, not a clever prompt. Rewriting the gag in the generator is a generate. You left start_template_run.

5. Edit a photo

Edit a photo with AI is generate_image_from_image. Required: prompt, image_url, models. Attach the image, describe the change, get an edit back. Cost is priced per edit model, shown before you confirm. There is no “edit this” without a this.

A text-to-image prompt that says “the same photo but the background is a beach” is a new picture of a beach, not an edit of yours. The starting file has to be on the wire.

Attach the hosted URL before the tool fires

The test is the attachment, not the sentence.

Job	Files on the wire	Tool
Animate one plate	One hosted image	generate_video_from_image + image_url
Bound the landing	Two hosted images	first + last frame
Copy a performance	Still + driving clip	motion-control mode
One-tap format	Template inputs as URLs	start_template_run
Change an existing jpeg	One hosted image	generate_image_from_image

Attach via the composer + button so the agent can read a Versely-hosted URL from chat history. Do not bury a path in the prompt and expect dispatch. Do not start the tool “to see if it asks.” The five capabilities exist because the catalog rows they call will not start empty.

Mixing jobs wastes the still

If you have no still, you wanted a text-only job. If you have a still and you typed a movie, you wanted a movie. If you have a still and a format name, you wanted one-tap. Mixing those sentences is how a prompt-only turn spends on a cousin object, or fails, and you still do not have the clip.

FAQ

Will the agent generate a still for me, then run these jobs?

Not as a silent first step of these five. Text-to-image is a different capability. These pages assume the plate already exists. If you needed a pack shot invented, generate the still on purpose, approve it, then attach that URL.

Is a URL I pasted in the prompt the same as an attachment?

No. The tools want a hosted URL the agent can pass as image_url, first_frame_url, last_frame_url, or template inputs. The + button is how that URL lands in chat. A sentence that mentions a file is still prompt-only.

Do all five use the same tool?

No. One-still video, first-last, and motion transfer share generate_video_from_image in different modes and different required fields. One-tap is start_template_run. Photo edit is generate_image_from_image. Same “needs a still” rule. Five jobs.

Can I pin GLM 5.3 so the agent “just knows” the photo?

No. z-ai/glm-5.3 is a text-only planner. It cannot take the image part. Pin x-ai/grok-4.6 if the still is the brief you want the brain to see. The generate tool still needs the hosted URL either way.