make_video

Generate a video — talking-head with native audio — from a prompt and optional reference images.

POST /api/v1/tools/make_video/run

Generates a video from a text prompt, optionally grounded in reference images. Speech is native: describe the person and write their lines in the prompt (e.g. She says: "...") and the model generates matching audio and lip movement. Optionally chains captioning onto the finished video in the same run.

Arguments

Argument	Type	Required	Default	Description
prompt	string	yes	—	What to generate, including any spoken lines. Max length is per model: 4000 chars for seedance-2, 2500 for kling-3. Over-cap prompts are rejected with a 400 — never silently truncated.
reference_images	array of strings (URLs)	no	—	Up to 14 publicly fetchable image URLs to ground the generation (a person, a product, a setting). Attaching any requires owns_references: true.
owns_references	boolean	with refs	—	Likeness-consent attestation — see below.
model	string	no	"seedance-2"	One of seedance-2, kling-3.
quality	string	no	"standard"	One of standard, pro. standard renders at 720p, pro at 1080p (each model’s premium tier). Pro costs more — see pricing.
duration	number	no	5	Video length in seconds. The server accepts 4, 5, 6, 8, 10, 12, 15 on seedance-2 and 3, 5, 8, 10, 15 on kling-3; other values are rejected with a 400. Cost scales with duration.
aspect_ratio	string	no	"9:16"	One of 16:9, 9:16, 1:1, 4:3, 3:4, 21:9.
audio	boolean	no	true	Generate native audio (speech, ambience). Leave on for talking-head content.
captions	boolean	no	false	When true, burns styled subtitles into the finished video within the same run, for an extra flat caption charge.
caption_style	string	no	"tiktok"	One of movie, tiktok. Only used when captions is true.
enhance_prompt	boolean	no	false	Rewrite a rough prompt into a production-grade one before generation — see below.
influencer	string	no	—	Slug of a saved influencer. The server attaches its portrait + character sheet as references and prepends its appearance description to your prompt. On seedance-2 this adds the standard +10% reference surcharge; on kling-3 the influencer takes 2 of the 4 reference slots (so at most 2 of your own reference_images alongside it). An unknown slug → 400, no debit.

owns_references consent

Any run that attaches reference_images must set owns_references: true — you attest that you own, or have the rights to use, the likenesses in the uploaded images. This applies to make_video and make_image alike.

• References attached without owns_references: true → 400, no credits debited.
• Accepted attestations are persisted on the run record (flag, timestamp, and the API key that made it) as an audit trail.

enhance_prompt

When true, your prompt is first rewritten by a prompt-writing model trained on what makes these video models perform — hook, framing, delivery — while preserving your intent and any spoken lines. The rewritten prompt is what gets generated; both the original and enhanced prompts are stored on the run (and are full-text searchable via GET /api/v1/generations?q=...).

Costs 3 credits extra. When false (the default), your prompt is sent to the model exactly as written.

Example — talking-head with captions

curl -X POST https://getagenthook.com/api/v1/tools/make_video/run \
  -H "Authorization: Bearer $AGENTHOOK_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "A man in his 30s in a gym, talking to camera, handheld selfie style. He says: \"Three exercises. Ten minutes. No equipment. Save this.\"",
    "quality": "standard",
    "duration": 5,
    "aspect_ratio": "9:16",
    "captions": true,
    "caption_style": "tiktok"
  }'

Example — with a reference image

curl -X POST https://getagenthook.com/api/v1/tools/make_video/run \
  -H "Authorization: Bearer $AGENTHOOK_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "The person from the reference image unboxes a pair of headphones at a kitchen table and says: \"Okay, these are lighter than they look.\"",
    "reference_images": ["https://example.com/my-face.jpg"],
    "owns_references": true
  }'

Response

202 with the queued run:

{
  "run_id": "run_abc123",
  "status": "queued",
  "credits_charged": 115
}

Poll GET /api/v1/runs/{run_id}; on completed, output[0] is the durable MP4 URL. On failed, credits are refunded instantly.