Roboflow Supervision

MIT-licensed Python CV toolkit — unified Detections type, model adapters, annotators, datasets, tracking, and metrics. "We write your reusable computer vision tools."

45,675+ GitHub stars (2026-06-30 API check); @skalskip92's milestone tweet cited 40K stars and 6.5K dependent OSS projects (7,714 bookmarks). Source: GitHub roboflow/supervision, 2026-06-30; Source: X/@skalskip92, 2026-06-06

Current Version Snapshot

Field Snapshot
Python package supervision==0.29.1
Python support >=3.9
Latest GitHub release 0.29.1, published 2026-06-23
Default branch develop, HEAD 934da124f5710fafcdb53e01596f0724ae68d8dd as of 2026-06-30
Repo status 45,675 stars / 4,055 forks as of 2026-06-30
License MIT

Source: PyPI JSON; GitHub API; git ls-remote, 2026-06-30

Core primitives

Module Purpose
sv.Detections Common box/mask container; converters from Ultralytics, Transformers, Detectron2, Roboflow Inference, YOLO-NAS, VLMs
Annotators Box/mask/label overlays on images/video
Datasets COCO, YOLO, Pascal VOC load/split/merge
Tracking & zones Persistent IDs, polygon zone counts
Metrics mAP, confusion matrices

Install: pip install supervision (Python ≥3.9). Docs: supervision.roboflow.com. Source: GitHub README, 2026-06-13

Minimal pattern

import supervision as sv
from ultralytics import YOLO

model = YOLO("yolov8n.pt")
results = model(source=0)[0]
detections = sv.Detections.from_ultralytics(results)
annotated = sv.BoxAnnotator().annotate(frame, detections)

Source: compiled from supervision.roboflow.com quickstart patterns, 2026-06-13

Ecosystem

Supervision is the shared utility layer in Roboflow's stack — not a trainer. Related repos: inference, autodistill, notebooks.

The basketball demo matters because it shows the shape of a real CV pipeline an agent can reason about: detector outputs, tracking IDs, segmentation overlays, team/color classification, jersey-number recognition, and shot/event classification become structured intermediate artifacts before an LLM ever sees them. For evals, Supervision is the layer that can turn video frames into measurable boxes, labels, zones, counts, and temporal tracks.

Stack fit (Kevin)

  • Agent vision demos — normalize model outputs before sending to LLM (count objects, zone triggers) in one Python layer.
  • Not on critical Dedalus path unless shipping CV features; prefer when building physical-world or video pipelines.
  • Pairs with Security and Review Skills for pulling demo videos, supervision for frame annotation in eval datasets.

Timeline