Call a saved workflow over HTTP as an endpoint.

Pick a workflow version and a GPU, set min and max workers, and get a URL. At 0 min workers the endpoint sleeps when idle, and the first request after that waits for a cold start.

Join the list Read the API docs

A request is a POST with the inputs you exposed on the workflow’s form.

request · brand-film · v7

curl https://api.milvey.example/v1/endpoints/brand-film/runs \
  -H "Authorization: Bearer $MILVEY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "input": {
      "product_image": "https://cdn.example.com/bottle.png",
      "prompt": "Slow orbit, soft rim light",
      "duration": 5
    },
    "webhook_url": "https://example.com/hooks/milvey"
  }'

response, then webhook

# 202 Accepted
{ "id": "req_91af", "status": "queued", "version": "v7" }

# POST https://example.com/hooks/milvey  (1m 29s later)
{
  "type": "run.succeeded",
  "id": "req_91af",
  "output": { "video_url": "https://…/req_91af.mp4" },
  "credits": 12,
  "gpu": "RTX PRO 6000"
}

Example request. Field names come from the workflow’s exposed inputs.

Min workers sets the trade between idle cost and cold starts.

At 0, an idle day costs 0 cr and the first request waits for a worker to load. At 1 or more, requests start at once and each awake worker is metered per second at its GPU’s cr/min rate.

Min workers 0Max workers 3Idle timeout 60s

0–3 workers, sleeps after 60 s idle. Idle cost 0 cr. Cold start ≈ 25 s.

Sample: RTX PRO 6000 at 8 cr/min while a worker is awake. Not a quote.

Each deploy phase is logged with its duration.

deploy · brand-film · v7 · RTX PRO 6000sample

00:00	Build: environment from the resolved node packs and weights	48 s
00:48	Push: image to the GPU region	21 s
01:09	Warm: weights loaded on the first worker	35 s
01:44	Health check: one test request against the input schema	6 s
01:50	Live at https://api.milvey.example/v1/endpoints/brand-film

A failed request shows 0 cr in the endpoint’s log.

Request	Time	Status	Duration	Credits
req_91af	12:41:07	200 · delivered	1m 29s	12 cr
req_91aecold start	12:38:55	200 · delivered	1m 58s	12 cr
req_91ad	12:30:12	422 · invalid input	0.2s	0 cr
req_91ac	12:22:40	200 · delivered	1m 31s	12 cr
req_91ab	12:10:03	500 · failed	44s	0 cr

Sample log for the brand-film endpoint.

An endpoint serves its pinned version until you move it.

Publishing a new workflow version leaves live endpoints alone.

brand-film

serving v7sample

Roll brand-film back to v5?

New requests run v5. Requests already running finish on v7.
v5 has no duration input. Calls that send it get a 422 and 0 cr.
Same URL, GPU and worker settings.

CancelRoll back to v5

Questions about endpoints

What does an idle endpoint cost?

With min workers at 0, nothing. An awake worker is metered per second at its GPU’s cr/min rate until the idle timeout puts it to sleep.

How do I get results for a long video?

The call returns a request id at once. The result is posted to your webhook_url, or you poll the id.

What happens when a request fails?

A failed request shows 0 cr in the log. How credits are reserved and returned is on the pricing page.

How are API keys handled?

A key is shown once when you create it. Each key has a scope, a last-used date and an optional spend cap. The snippets read it from an environment variable.

Can I change the GPU later?

Yes. Changing the GPU or worker counts redeploys the endpoint on the same URL and the same version.

Send me the workflow your app needs to call.

Join the list and reply with the graph. I’ll deploy it as an endpoint and send you the curl command and the credits per request.

Email when signups open · Unsubscribe any time