Bio Model v0 is in development. No results yet, and we'll publish the misses with the hits. Join the waitlist

Developers

One call for every animal in the frame. Send video or a live stream, get typed JSON back. The API is in development, so this is the plan.

analyze.pyPython
import vitals

# a file, a URL or an RTSP stream
result = vitals.biomodel.analyze(
    "barn-cam-2.mp4", species="cattle"
)

for animal in result.animals:
    if animal.state.label != "calm":
        print(animal.id, animal.state.label, animal.state.side)
    path = animal.next.path        # where it's likely to be, next 3 s

# live: one event per animal per frame
for event in vitals.biomodel.stream("rtsp://10.0.0.5/live"):
    robot.avoid(event.next.path)
analyze.tsTypeScript
import { biomodel } from "@vitals/sdk";

const result = await biomodel.analyze(video, {
  species: "auto",
});

for (const a of result.animals) {
  if (a.next.intent === "about to bolt") slowDown();
}

The response. Typed, the same shape for every species, one entry per animal per frame.

Every field is something a machine can act on without a person in the loop: where the animal is, how it's standing, how it's doing, and where it's going.

Planned shape. The values here are made up to show the format; no trained model produced them.

response.jsonPlanned
{
  "frame": 1842,
  "time": "2026-11-19T06:14:03.2Z",
  "animals": [{
    "id": "C-0412",
    "species": "cattle",
    "box": [412, 188, 655, 470],
    "pose": { "keypoints": [[598, 214, 0.97], ...] },
    "motion": { "speed_mps": 0.6, "heading_deg": 92 },
    "state": {
      "label": "lame", "side": "left hind",
      "vs_own_normal": "off for 2 days", "confidence": 0.71
    },
    "next": {
      "horizon_s": 3,
      "path": [[640, 470], [688, 468], [731, 466]],
      "intent": "walking to water"
    }
  }]
}

Fields. What each part of the response means.

idA stable ID for the individual, kept across frames and cameras.
speciesCattle, horse, pig, sheep, goat, chicken, dog, cat, deer, or "other".
box, poseWhere it is in the frame, and keypoints for head, spine and legs, each with a visibility score.
motionSpeed, heading and gait.
stateCalm, stressed, in pain, lame or unwell, against the animal's own normal, with a confidence.
nextA short forecast: the likely path over the next few seconds and what the animal looks to be doing.

How it will run. Where the video is.

On the device

A small runtime for robot and camera boards, so a milking arm or a drone gets answers without a round trip. Video stays on the device.

In the cloud

A REST API for files and streams, with signed webhooks when an animal's state changes.

SDKs

Python and TypeScript first. Tell us what your robot runs when you join the waitlist.

None of this is released. Bio Model v0 is being trained now, and we'll publish what its locked tests show, misses included, before anyone builds on it.

Build on the first model of animals.

Bio Model v0 is in development. Join the developer waitlist for early API access, or pilot it on your own cameras.

One call, typed JSON

Send video, get back each animal's identity, pose, state and next few seconds.

API and SDK
See the first app on it

Vitals Cam tells a rancher which animal to check today.

Vitals Cam