Spotr

Runs in the browser · no account needed to try

Describe what you want to find.
Get a detector that finds it.

Spotr takes a sentence and a list of classes, generates a synthetic dataset, labels it with bounding boxes, trains a YOLO detector on a free GPU, and hands you a link. Open the link on any phone or laptop and detection runs entirely on-device — your camera frames never leave the browser.

Try a live demoModel creation opens up next — the demos below are real and work right now.

Try a demo

All three demos run one COCO-pretrained YOLO26n graph, filtered to a focused class list. Models trained by Spotr replace them once the training pipeline lands.

How it works

  1. 01

    Name your classes

    Give the model a name and the things it should find — “hard hat, safety vest”.

  2. 02

    Generate or upload

    A language model expands your scene into ~40 varied prompts, or bring your own photos.

  3. 03

    Review the labels

    Boxes are drawn automatically by an open-vocabulary detector. Delete the bad ones before training.

  4. 04

    Share the link

    Training exports to ONNX and you get a six-character URL that runs on any modern browser.

Built to cost nothing to run

Every part of Spotr sits on a free tier, which is a design constraint rather than a footnote. Inference runs on your device instead of a server, so a popular model page costs the same as an unpopular one. Training runs on donated serverless GPU credits with a hard per-job time cap, and falls back to a second free backend when the monthly budget runs out.

Inference
On-device, WebGPU with a WASM fallback
Detector
Ultralytics YOLO26 nano at 640px
Training cap
10 minutes per model
Frames uploaded
None, ever
Licence
AGPL-3.0, source on GitHub
Cost to you
Free