Open-source alternatives to Labelbox: 3 data labelling and annotation tools compared
Labelbox holds the data, puts it in front of annotators and returns labels in its own format; the three open-source candidates replace the platform but not the workforce, so the annotator contracts and the data handling stay yours. HumanSignal/label-studio is the general-purpose pick, cvat-ai/cvat for vision work, argilla-io/argilla for LLM feedback.
What Labelbox does, and why people look elsewhere
Labelbox is a hosted annotation platform: it holds the images or text to be labelled, sends them to annotators, and returns the labels together with any model-assisted pre-labels.
- the data being labelled and the resulting labels are stored on the vendor's servers
- annotators work in the vendor's editor, so the label format is the vendor's
- there is no self-hosted version of the same workflow, so annotation needs an outside service
The candidates at a glance
| Project | Licence | Maintenance | Language | GitHub stars | Install |
|---|---|---|---|---|---|
| HumanSignal/label-studio | Apache-2.0Permissive: commercial use allowed | Commits in the last dayLast push October 23, 2026 | TypeScript | 28,540 | See analysis |
| argilla-io/argilla | Apache-2.0Permissive: commercial use allowed | Commits in the last six monthsLast push October 19, 2026 | Python | 5,175 | See analysis |
| cvat-ai/cvat | MITPermissive: commercial use allowed | Commits in the last six monthsLast push October 9, 2026 | Python | 16,899 | See analysis |
Each alternative in detail
HumanSignal/label-studioLabel Studio is a multi-type data labeling and annotation tool with standardized output format
- Best for
- A team that labels a mix of media types, because one instance covers audio, text, images, video and time series behind a single interface and exports to model formats from the same tool.
- What you give up
- It is a server you run, so the database choice is yours: the Compose stack replaces SQLite with PostgreSQL and puts Nginx in front of it, and adding MinIO for object storage means giving the container a hostname on your network. The Apache-2.0 licence covers the code in this repository rather than the hosted editions, and the label shape your pipeline consumes has to be confirmed against the export the instance produces rather than assumed.
argilla-io/argillaArgilla is a collaboration tool for AI engineers and domain experts to build high-quality datasets.
- Best for
- An LLM team collecting human feedback on model outputs, where records are filtered, ranked and revised in a browser and the result feeds straight back into preference tuning or evaluation.
- What you give up
- It is built for feedback on text and model output rather than for drawing on images, so a vision pipeline sits outside what it was designed for. The README states that the original authors have moved on and that no new features are planned, with bug fixes and patches as needed, so a roadmap-driven adoption is a poor fit, and the quickest route to a running instance is a hosted space template rather than a deployment you own.
cvat-ai/cvatComputer Vision Annotation Tool (CVAT) is a leading platform for building high-quality visual datasets for vision AI. It offers open-source, cloud, and enterprise products, as well as labeling services, for image, video, and 3D annotation with AI-assisted labeling, quality assurance, team collaboration, analytics, and developer APIs.
- Best for
- A computer vision team that needs image, video and 3D annotation with team roles, task assignment and a review workflow, and that needs the media to stay inside its own infrastructure.
- What you give up
- The MIT licence covers the community build, and the README warns that some serverless assets and dependencies may carry separate licences, so what actually runs has to be checked. The stack is heavier than the other two, with Docker Compose, a database and object storage to run, the clients are primarily tested on Chromium-based browsers, and single sign-on, the built-in SAM auto-labelling and AI agents sit outside the community code.
- Moving from Labelbox
- CVAT imports and exports in documented interchange formats, including COCO, YOLO, Pascal VOC and KITTI, so an existing labelled dataset can be brought in through a format instead of being relabelled from scratch.
Three ways to run annotation, and the part none of them replaces
Labelbox does three separable things. It holds the data to be labelled, it gets that data in front of annotators, and it returns labels, sometimes pre-filled by a model. The three candidates here cover the first and the third well. The second is the one they do not.
HumanSignal/label-studio is the general-purpose tool: one interface, one standardised export, and support for audio, text, images, video and time series in the same instance. It is a server you run, and the annotators are whoever you point it at.
argilla-io/argilla is narrower and aimed at a different kind of data. It is a collaboration surface for AI engineers and domain experts working on language and model datasets, where the unit of work is a response to a prompt, a suggestion from a model, and a human judgement between the two. Filters, semantic search and AI feedback suggestions are the tools it provides.
cvat-ai/cvat is the vision one. It covers image, video and 3D annotation with drawing tools, roles and task assignment, a review workflow, and an API for the actions that need scripting. It also connects to your own models for detection, segmentation and tracking, which is the pre-label step.
What none of them does is supply annotators. Labelbox has a workforce and a review process around it. The three open-source projects expect you to find people, contract them, and handle their access to the media yourself. That is the largest difference between the product and its replacements, and it is not a deployment detail. It is why the annotation contract, the data processing terms and the access controls become your problem in a way they were not before.
Where the label format lives, and how pre-labels arrive
Labelbox returns labels in the format the product defines, and annotators work in the vendor's editor, so the schema sits on the vendor's side of the bargain. A self-hosted tool means owning the schema, and the candidates differ a lot in how much help they give.
cvat-ai/cvat is the strongest on interchange. It imports and exports in documented formats, including COCO, YOLO, Pascal VOC and KITTI, which means a dataset labelled elsewhere can come in through a format instead of being relabelled by hand. That matters more than it sounds, because re-drawing a large corpus is the cost that quietly decides whether a migration is worth starting.
HumanSignal/label-studio promises one interface and one standardised export across media types, which is the right shape for a team that does not want five converters. The export is a versioned format of its own, so the useful question is what your training pipeline consumes, checked against the export the instance actually produces rather than assumed.
argilla-io/argilla is not trying to be a general annotation format. Its records are feedback on model output, which is what makes it a good fit for preference tuning and evaluation and a poor substitute for a labelled image set.
Pre-labels are the other half of the workflow. cvat-ai/cvat connects to your own models for detection, segmentation and tracking, the closest match to a model-assisted pre-label pipeline. Label Studio can call machine learning models to pre-annotate, and Argilla offers AI feedback suggestions inside the review. None of them runs a proprietary model for you, so you bring the model or you do without.
What each one costs to run
Label Studio has the gentlest start. A single container with a data directory keeps the SQLite database and the uploaded files, which is enough to try the tool. The Compose stack replaces SQLite with PostgreSQL and puts Nginx in front of it for static assets, and a second Compose file adds MinIO so you can test object storage locally, which needs a hostname entry on your network if the host has no static address. Beyond that there are pip, poetry and Anaconda routes, and a development mode that runs the migrations by hand. The database is worth choosing up front, since SQLite and PostgreSQL are different stores.
argilla-io/argilla is the smallest of the three to get going: a Python package, with a space template as the quickest route to a running instance and a sign-in through a Hugging Face account for the browser interface. Hosting it yourself is a Python deployment with the usual database and storage decisions behind it.
cvat-ai/cvat is the heaviest, and the README is upfront about the prerequisites: a container runtime, Docker Compose and Git. You clone the repository, bring the stack up, and create an administrator account before the first session. The deployment guides cover AWS, Kubernetes, an external PostgreSQL, backups and upgrades, which tells you the shape of the work. The browser side has its own constraint: the clients are primarily tested on Chromium-based browsers, Firefox works with caveats, and Safari and WebKit are not supported, so a team on Macs should hear that before annotators are given the link.
Licences, and the assets underneath them
Two of the three are Apache-2.0, HumanSignal/label-studio and argilla-io/argilla, and one is MIT, cvat-ai/cvat. All three are permissive in the ordinary sense: you can run them inside a company, change them, and host the modified version without an obligation to publish your changes. None of them is copyleft, so hosting a modified instance over a network attaches no source-publication condition, and that part of the picture settles quickly.
The conditions that remain are worth reading rather than assuming. The Apache-2.0 grant on HumanSignal/label-studio covers the code in that repository, and the README points readers to the hosted editions separately, so the terms covering the hosted product are a different matter from the terms covering the code you would deploy. cvat-ai/cvat carries a caveat of the opposite shape: the MIT licence covers the community build, and the README warns that some serverless assets and dependencies may come with separate licences, which is the sort of detail that turns up in a procurement review.
Then there is the licensing that does not apply to code at all: the contracts with the people doing the labelling, the data processing terms, and where the media sits while it is being labelled. Under a hosted platform those sit with the vendor. Once you run the tool yourself they are yours, and for an annotation project that is usually a bigger question than the choice between Apache-2.0 and MIT.
Choosing by scenario, and where each one is the weaker choice
For a team labelling a mixed bag of media, HumanSignal/label-studio is the pick, because one instance and one export format covers audio, text, images, video and time series, and the Compose stack with PostgreSQL and Nginx is a familiar shape to operate.
For a computer vision team building training sets for detection, segmentation or tracking, cvat-ai/cvat is the pick, because the drawing tools, the roles, the review workflow and the interchange formats are built for that work, and the media stays on your own infrastructure. Accept the Chromium constraint and the heavier stack.
For an LLM team collecting preference data and human feedback, argilla-io/argilla is the pick, with the understanding that the README says no new features are planned, so the roadmap will not help you and a missing capability will not be added for you.
The weaker choice in each pairing is worth naming. Label Studio is the weaker option for a vision pipeline that needs CVAT's drawing and tracking features, since its own media list is audio, text, images, video and time series. cvat-ai/cvat is the weaker option for text and audio, which are outside what it annotates. argilla-io/argilla is the weaker option whenever the deliverable is an image with a box on it, because its workflow is built around records and judgements rather than drawings. And all three are the weaker option when the annotator workforce was the point, because none of them supplies people.
Bottom line
Take HumanSignal/label-studio for mixed media, cvat-ai/cvat for vision and 3D work that must stay on your own servers, and argilla-io/argilla for feedback on model output. Before committing, export a small labelled set from each candidate and check the format your pipeline reads, and confirm the browser, database and object storage arrangements your annotators will meet on day one.
Sources
Was this useful?