CVAT
MIT-licensed annotation tool for images, video and 3D point clouds. Self-host via Docker, plug SAM or YOLO into Nuclio for pre-labelling, export to COCO, YOLO, KITTI and 20 other formats.
Overview
Started inside Intel as tooling for the OpenVINO team, open sourced in 2018 and now run by CVAT.ai, with the self-hosted Community edition under MIT. Version 2.74.0 shipped on 26 August 2026, which says most of what needs saying about the pace. Images, video and 3D point clouds are all first-class citizens, annotated with boxes, polygons, masks, keypoints, cuboids and skeletons.
Auto-annotation runs through Nuclio serverless functions. You deploy prebuilt functions for Segment Anything, the SAM 2 tracker or YOLO into your own cluster, then call them from the canvas so annotators correct rather than draw from scratch. Nothing leaves your network, which is the whole argument for self-hosting in the first place. The trade is operational overhead: multiple Compose files, a version-matched nuctl and Redis if you want SAM 2 tracking state to persist between calls. Export covers more than twenty formats, so COCO for detection and KITTI for anything driving-shaped come out without a conversion script.
CVAT Online is the hosted route. Solo runs $33 a month, $23 on annual commitment, Team is the same per-seat rate for two to fifty members and Enterprise starts near $12,000 a year for on-prem with SSO and support SLAs. For a vision team sitting next to the production line the calculation is usually short: run Community on hardware you already own and put the budget into GPUs instead.
Key Features
- ✓ Boxes, polygons, masks, keypoints, cuboids and skeletons across images, video and 3D point clouds
- ✓ Serverless pre-labelling through Nuclio with prebuilt Segment Anything, SAM 2 tracker and YOLO functions running inside your own network
- ✓ Frame interpolation that carries a shape across a long clip instead of redrawing per frame
- ✓ 20+ import and export formats including COCO, YOLO, Pascal VOC and KITTI
- ✓ Hierarchical quality requirements at task and project level, added in v2.74.0
- ✓ Python SDK and CLI with profile-based auth and personal access tokens
Where it holds
- • MIT on the self-hosted edition, so no seat arithmetic and no vendor sitting between you and your own images
- • Video interpolation is the strongest reason to pick it over Label Studio for clip-heavy work
- • 3D point cloud and cuboid support is rare in free tools and directly useful for bin-picking datasets
- • Release cadence is roughly fortnightly: v2.74.0 landed 26 August 2026
Where it breaks
- • The Compose stack is heavy. Auto-annotation needs a second and third compose file, and the nuctl binary must match the version pinned in docker-compose.serverless.yml (1.13.0 at time of writing)
- • The Nuclio dashboard on port 8070 ships with no authentication, so binding it to 0.0.0.0 hands anyone on the network a way to deploy arbitrary functions
- • CVAT Online's free tier is a demo, not a working allowance: 1 project, 3 tasks, 1 GB
- • The review and QA workflow is thinner than what Label Studio Enterprise offers, and the UI slows on very large tasks
My Take
Video is where it earns its keep. Interpolating a box across four hundred frames and fixing the drift beats redrawing, and the cuboid tools cover point cloud work that Label Studio still handles badly. Setup is the tax: the auto-annotation side wants an extra Compose file plus a version-matched nuctl, and the Nuclio dashboard on 8070 has no auth at all, so bind it to localhost or you have quietly published a remote execution endpoint. Worth the trouble for an MIT licence and a fortnightly release train.
Quick Info
- Pricing:
- open-source
- Openness:
- Open source
- Licence:
- MIT
- Starting at:
- Self-hosted Community edition is MIT and free with no seat limit. CVAT Online free tier caps at 1 project, 3 tasks, 1 GB storage and 100 AI tool calls a month. Solo is $33/month or $23/month billed annually. Team is the same $33 per user per month ($23 annually) for 2 to 50 members, with internal storage scaling 25 GB to 250 GB and AI calls 10,000 to 100,000 a month by tier. Enterprise, meaning your own cloud or on-prem with SSO and SLAs, starts around $12,000 a year. Verified on cvat.ai pricing pages, August 2026.
- Added:
- Aug 2026
- Updated:
- Aug 2026
Use Cases
Judge it on your own work
The notes above say where CVAT holds and where it breaks. The fastest check is your own workload.
Visit website ↗Alternatives to CVAT
Label Studio
open-sourceApache 2.0 labelling tool covering images, audio, text, video and time series in one config-driven UI. Self-host free, or $99 a month for HumanSignal's hosted Starter tier.
Roboflow
freemiumEnd-to-end computer vision platform for labelling, training and deploying detection models to cloud or edge hardware like Jetson, with a visual builder for multi-step vision pipelines.
FiftyOne
open-sourceApache 2.0 Python library and app for curating vision datasets. Query and slice samples, explore embeddings, find label mistakes and push fixes back to CVAT or Label Studio.