Best Open-Source Data Labeling Tools in 2026
By Daniel Clarke
If you need to label training data without vendor lock-in or recurring licence costs, open-source annotation tools are the obvious starting point. The field is smaller than the listicles suggest: two tools dominate, and most of the rest are either abandoned or not actually open source.
This page covers both, plus the near-misses worth knowing about. Two things here that you won't find on the vendor blogs ranking above us: real GitHub maintenance data, and an honest account of which features are paywalled in each "open source" tier — the single most important fact if you're planning to self-host and never pay.
GitHub data pulled 2026-07-17. Pricing and tier limits last checked 2026-07-17.
The short answer
CVAT for computer vision — images, video, 3D/LiDAR. Label Studio for everything else — text, audio, time series, documents, LLM evaluation.
That single distinction resolves most decisions. If you want the long-form head-to-head, we wrote one: CVAT vs Label Studio.
Comparison matrix
| Tool | Data types | Self-host | Licence | Stars | Paid-tier gating | Best for |
|---|---|---|---|---|---|---|
| CVAT | Images, video, 3D/LiDAR | Free | MIT | 16,317 | None — self-hosted is the full product | Computer vision |
| Label Studio | Images, video, text, audio, time series, PDF | Free | Apache-2.0 | 27,855 | RBAC, SSO, QA workflows, analytics | Multimodal, LLM eval |
| Supervisely | Images, video, 3D, DICOM | Free tier | Not fully OSS | — | Most enterprise features | Medical imaging |
| Roboflow | Images, video | No | Proprietary | — | Private data requires paying | Fast CV projects |
CVAT: best for computer vision
CVAT is the most widely adopted open-source image annotation tool, backed by the OpenCV Foundation and originally built at Intel. It holds a 4.8/5 rating on G2 — the highest of any open-source annotation tool.
Strengths:
- Native 3D point cloud and LiDAR annotation — genuinely rare in open source
- 19 export formats including YOLO, COCO, PASCAL VOC, and KITTI
- Auto-annotation with AI models, SAM 2/3 integration, claims up to 10x faster labeling
- Cloud storage integration for S3, GCP, and Azure
Limitations:
- Computer vision only — no text, audio, or document support
- Performance degrades with very large video files or thousands of images
- Self-hosted deployment requires real technical expertise
Gating: none that matters. Self-hosted CVAT is the complete product under MIT. This is the strongest "free actually means free" story in the category.
Pricing: free to self-host. Cloud from $23/mo; Enterprise around $12,000/year. The free cloud tier (1 project, 3 tasks, 1GB) is an evaluation sandbox, not a workspace.
Choose CVAT if: you're building computer vision models and want the deepest export support without enterprise pricing.
Label Studio: best for multimodal data
Label Studio is the most flexible open-source data annotation tool, handling images, video, text, audio, time series, and PDFs in one place. It has 1M+ users and a 20K+ member Slack community — the largest community in open-source annotation.
Strengths:
- Broadest data type support of any open-source tool
- Highly customizable XML-based labeling templates — you can build almost any interface
- LLM evaluation, RLHF, and RAG validation workflows
- Strong Python SDK; installs via pip, brew, or Docker
Limitations:
- Steep learning curve — the XML templates take real time to master
- Performance issues with very large datasets
- Community Edition has no role-based access control
Gating — read this before committing: Label Studio Community includes all core labeling, the API/SDK, and import/export. But RBAC, SSO, reviewer assignment, QA/review workflows, advanced analytics, and SOC 2 are Enterprise-only. If you have a team of annotators who need permissions and a review stage, Community Edition will not do it, and the fix is a sales conversation. A solo researcher won't care. A 20-annotator operation will hit this in week one.
Pricing: free to self-host. Starter Cloud $50/mo; Enterprise custom.
Choose Label Studio if: you need multiple data types in one tool, or you're doing LLM evaluation.
Are they actually maintained?
Nobody in this category publishes maintenance signals, which is odd, because for open-source software it's the first thing you should check. From the GitHub API on 2026-07-17:
| Signal | CVAT | Label Studio |
|---|---|---|
| Stars | 16,317 | 27,855 |
| Forks | 3,764 | 3,624 |
| Open issues | 601 | 900 |
| First commit | June 2018 | June 2019 |
| Latest commit | 2026-07-16 | 2026-07-17 |
| Commits in last 90 days | 100+ | 100+ |
| Latest stable release | v2.70.0 (2026-07-14) | 1.23.0 (2026-03-13) |
| Release cadence | ~every 2–4 weeks | ~quarterly + nightlies |
Both are healthy. Each exceeded 100 commits in the trailing 90 days (that's the API's page cap, so both are higher) and both were committed to within a day of this update.
The difference is release philosophy. CVAT tags small releases every few weeks. Label Studio ships quarterly stable releases with a rolling nightly in between — so the four-month gap since 1.23.0 is cadence, not neglect. If you self-host and want fixes in tagged releases rather than nightlies, CVAT delivers faster.
Other options worth considering
Supervisely has a free Community tier with built-in AI models (SAM2, YOLO v11) and strong DICOM support. It is not fully open source — the free tier is functional for researchers and small projects, but the platform is a commercial product with most enterprise capability behind paid tiers. Its LiDAR and medical imaging support is genuinely good.
Roboflow isn't open source either, but its free Public tier includes $60/month in credits, 2 users, and full annotation, training, and deployment. The catch is significant: your data and models become open source on Roboflow Universe. For research and learning that's often fine. For anything proprietary, the Core plan starts at $79/month. Worth being explicit about, because "free tier" and "your dataset is public" are very different propositions.
What "free" actually costs
Both CVAT and Label Studio self-host via Docker and hand you full control of your data. Neither is free to run. Self-hosting requires:
- Server infrastructure (cloud VM or on-premise) with enough RAM to survive your largest files
- Engineering time for upgrades, backups, storage, and auth
- Your own security posture
That's the real trade: a vendor invoice for engineering hours. With DevOps capacity it's usually a good deal — often dramatically cheaper than per-seat enterprise pricing. Without it, the $23–50/mo managed tiers are cheaper than they look once you count the hours.
One trap worth naming: the performance ceiling is not a hosting misconfiguration. Both tools have documented slowdowns on very large datasets, and no amount of Docker tuning fixes it. Size your hardware for your actual data.
Bottom line
CVAT for pure computer vision with advanced export needs and no feature gating. Label Studio for multimodal data or LLM workflows — as long as you either don't need team permissions or are willing to pay for them.
Both are production-ready and genuinely free to self-host. The honest caveat is that "open source" means something meaningfully different in each: CVAT gives you everything, Label Studio holds back the team-management layer.
If you're weighing these two specifically, see CVAT vs Label Studio. If you're evaluating commercial platforms instead, see our Scale AI alternatives comparison, or the Appen vs Scale AI head-to-head.