AI video generation API for apps and agents

Three calls: quote, start, wait

Video runs take one to several minutes, so the API never holds a POST open for that long. Instead, POST /api/v1/videos charges the run and answers at once with a job in queued. You then read GET /api/v1/jobs/{id}?wait=50, which holds the line for up to 50 seconds and returns early the moment the job is final.

Before the POST, send the same body to POST /api/v1/quote. It is free, it charges nothing, and the number it returns is what the run will cost.

const KEY = process.env.KINOSEED_API_KEY!; // ks_live_YOUR_KEY
const BASE = "https://kinoseed.app/api/v1";
const headers = { Authorization: `Bearer ${KEY}`, "Content-Type": "application/json" };
const body = {
  prompt: "a lighthouse at dusk, waves breaking, slow aerial orbit",
  model: "kling-3",
  duration: 5,
  aspect: "16:9",
};

// 1. Price it. Free, and the number is what the run will charge.
const quote = await (
  await fetch(`${BASE}/quote`, { method: "POST", headers, body: JSON.stringify({ kind: "video", ...body }) })
).json();
console.log("this clip costs", quote.credits, "credits");

// 2. Start it. The Idempotency-Key makes a retry after a timeout safe.
const res = await fetch(`${BASE}/videos`, {
  method: "POST",
  headers: { ...headers, "Idempotency-Key": "lighthouse-001" },
  body: JSON.stringify(body),
});
if (!res.ok) throw new Error((await res.json()).error.code); // e.g. insufficient_credits
let job = await res.json();

// 3. Wait for it. Each GET holds up to 50 seconds and returns as soon as the job is final.
while (job.status !== "completed" && job.status !== "failed") {
  job = await (await fetch(`${BASE}/jobs/${job.id}?wait=50`, { headers })).json();
}
console.log(job.status, job.output.videoUrl ?? job.error, job.credits); // a failed job shows its refund

Each step is documented on its own page: Quote, Videos and Jobs.

The models and what a clip costs

Every model below is reachable with the same request shape; only the model id changes. Prices are for the shortest clip each model makes, at its first resolution, without sound, which is the example GET /models returns:

• seedance-2 Seedance 2.0: 130 credits ($2.60) for 5s, at 480p.
• kling-3 Kling 3.0: 110 credits ($2.20) for 5s, no sound.
• veo-3-1 Google Veo 3.1: 150 credits ($3) for 4s, at 720p, no sound.
• veo-3-1-lite Google Veo 3.1 Lite: 25 credits ($0.50) for 4s, at 720p, no sound, from a start frame.
• grok-imagine-video Grok Imagine: 55 credits ($1.10) for 6s, at 480p.
• wan-3 Wan 3.0: 50 credits ($1) for 5s, at 480p.
• minimax-h3 MiniMax H3: 50 credits ($1) for 5s, at 480P.
• seedance-2-5 Seedance 2.5: 170 credits ($3.40) for 4s, at 480p.
• wan-2-7 Wan 2.7: 95 credits ($1.90) for 5s, at 720p, from a start frame.
• minimax-hailuo-23 MiniMax Hailuo 2.3: 90 credits ($1.80) for 6s, from a start frame.
• seedance-1-5-pro Seedance 1.5 Pro: 12 credits ($0.24) for 5s, at 480p, no sound.
• seedance-2-mini Seedance 2.0 Mini: 70 credits ($1.40) for 5s, at 480p.
• kling-3-standard Kling 3.0 Standard: 80 credits ($1.60) for 5s, no sound.
• seedance-2-fast Seedance 2.0 Fast: 110 credits ($2.20) for 5s, at 480p.
• kling-2-6 Kling 2.6 Pro: 65 credits ($1.30) for 5s, no sound.
• minimax-hailuo-23-fast MiniMax Hailuo 2.3 Fast: 40 credits ($0.80) for 6s, from a start frame.
• veo-3-1-fast Google Veo 3.1 Fast: 75 credits ($1.50) for 4s, at 720p, no sound.

Length, resolution and sound change the price. A 5-second Kling 3.0 clip is 110 credits; a 10-second one with sound is 310. The text to video API pricing guide explains what moves the number, and the video generation API comparison puts each model next to the provider's list price.

Text, frames and references

Models differ in what they accept, and GET /models says so per model in its inputs field. The request fields are the same for all of them:

• prompt: what happens in the clip. Up to 2,000 characters.
• startImageUrl: a first frame to animate. Models that only animate images need it; text models switch to their image-to-video endpoint when you send it.
• endImageUrl: a last frame, on the models whose inputs.endImage is true.
• referenceImageUrls: several images the prompt refers to, on models with a multi-reference mode.
• duration, aspect, resolution: snapped to the nearest value the model allows, and the quote follows the snapped value.
• audio: off by default, because sound can double the price on some models. Turn it on when you want it.

Input images must be public https URLs or URLs returned by POST /files. The API resolves every URL and refuses private addresses, localhost and plain http before anything fetches it, so a key holder cannot point it at an internal network.

When a run fails

Providers fail some runs: a prompt is refused, a frame cannot be read, a queue times out. When that happens the whole charge comes back on its own and the job shows it: credits.charged and credits.refunded are equal, and error says why in plain words. There is nothing to claim.

Retries are safe if you send an Idempotency-Key. The same key with the same body within 24 hours returns the first job instead of starting and charging a second one. See Idempotency.

An account can have 5 runs queued or running at once, web app runs included. One more is refused with 429 too_many_running and nothing is charged.

Polling or webhooks

Polling with ?wait=50 is enough for a script or an agent. A backend with many runs can register a webhook and Kinoseed sends a signed job.completed or job.failed POST with the whole job in it. The output URLs are already copied to Kinoseed's storage, so they do not expire with the provider's.

curl -X POST https://kinoseed.app/api/v1/webhooks \
  -H "Authorization: Bearer ks_live_YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{ "url": "https://example.com/hooks/kinoseed", "events": ["job.completed", "job.failed"] }'
# The response holds "secret": "whsec_..." once. Store it now.

See Webhooks.

From an agent

The same runs are available as MCP tools: generate_video starts a run and returns the job, and get_job with wait_seconds follows it. Claude Code, Cursor and other MCP clients connect with one URL and the same API key. The Claude Code guide walks through it, and the MCP server page has the setup for every client.

Generating begins with a plan from $10 a month (500 credits a month). Every tool page is open to browse without an account.