Guide

Self-hosted image annotation tools: what to check before you install one

Five image annotation tools you can run on your own servers, compared from their own documentation. What the free editions include, what sits in the paid ones, and what running it yourself involves.

Contents

  1. Two different searches under one name
  2. Before you install anything, decide what has to stay inside
  3. What to check
  4. Who operates it
  5. Accounts, roles and sign-in
  6. Review and a record of changes
  7. Where files are stored
  8. Updates
  9. What is free and what is not
  10. The tools, side by side
  11. CVAT
  12. Label Studio
  13. VIA
  14. BIIGLE
  15. DataTorch
  16. A short way to decide
  17. Sources
Contents

Two different searches under one name

People who search for a self-hosted annotation tool usually mean one of two things.

The first is a free, open-source tool to run on a workstation or a lab server. That exists, it is well documented, and for one person or a small group it can be working in an afternoon.

The second is an annotation system inside a company's network, with named accounts, sign-in through the company's identity provider, review, and a record of who changed what. That also exists. At every vendor we read, ours included, it is a paid edition and it starts with a conversation.

This page covers both, from each tool's own documentation as read on November 17, 2026. Where a page did not say something, the table says so.

Three ways to keep annotation data under your control: a hosted tool with hosted storage, a hosted tool reading your own bucket, and the tool running inside your network
Three routes, by where the files live and who runs the tool. Diagram by Lumen.

Before you install anything, decide what has to stay inside

Self-hosting is a means. The requirement behind it is usually one of these three, and they lead to different answers.

  1. The files must stay in storage we own. A hosted tool that reads from your bucket can meet this. Nobody on your side operates a server.
  2. Nothing may leave our network. The tool itself has to run inside. Any of the tools below can do this.
  3. No internet connection at all. An air-gapped install. Check how the tool is installed and updated without reaching a package registry, and whether its model-assisted features still work offline.

If your requirement is the first one, read the storage section of any hosted tool before taking on a server.

What to check

Who operates it

A self-hosted tool is a service someone on your team runs: a database, file storage, backups, certificates and upgrades. CVAT's upgrade guide is a fair picture of the routine. It says "It is highly recommended backup all CVAT data before updating", then stop the containers, get the new source, start again, and "do not terminate the migration and wait till the process is complete". None of that is hard. It does need an owner.

Accounts, roles and sign-in

A tool on one machine for one person needs no accounts. A team does, and a regulated team usually has to use its company sign-in. This is the line most free editions stop at. See the table.

Review and a record of changes

If someone will ask later who labelled an image, who changed it and who approved it, check for reviewer assignment and an activity or audit log before you choose, and check which edition they are in.

Where files are stored

Check whether the tool copies your images into its own volume or reads them in place from storage you already have, and what a backup then has to include.

Updates

Check how a new version arrives, how often, and what happens to the database when it does. On an air-gapped network, check how it arrives at all.

What is free and what is not

Read the vendor's own comparison table. The free editions below are real products, not trials. The paid editions add specific things, and they are listed.

The tools, side by side

CVATLabel StudioVIABIIGLEDataTorch
License of the self-hosted editionMIT (CVAT Community)Apache 2.0 (Community)BSD 2-clauseMITCommercial, Enterprise plan
How it installsDocker Compose, one command after cloning the sourcepip, Docker, or Docker Compose with PostgreSQLNone. A single HTML file opened in a browserDocker Compose after building the images, on UbuntuHelm chart on Kubernetes, in your cloud or on-prem
Accounts in the free editionYes. A superuser is created from the command lineYes. Sign up and log in, annotations tied to the account. No role-based access controlNo accountsNot covered in the pages we readNo free self-hosted edition. Free and Team are hosted
Company sign-in (SSO)CVAT Enterprise and paid online plansEnterprise (SAML SSO or LDAP)NoNot covered in the pages we readAsk us. Not stated on our public pages
Reviewer assignment and activity logQuality control and analytics are listed under paid plansReviewer assignment in Starter Cloud and Enterprise. Activity logs in EnterpriseNoNot covered in the pages we readReview file by file, dataset sign-off as a locked version, audit trail
Air-gappedNot covered in the pages we readNot covered in the pages we readRuns offline by designNot covered in the pages we readSupported, with an offline license
Paid editionEnterprise Basic and Premium, by sales contactStarter Cloud, and Enterprise with on-premise availableNoneNone found. A hosted instance runs at biigle.deEnterprise, priced per account

CVAT

CVAT Community is, in its own words, "the free, self-hosted open-source edition of CVAT", released under the MIT license. Its README says: "Run entirely within your own infrastructure. No data leaves your environment."

The install guide covers Ubuntu, Windows 10 through WSL2 and macOS, and comes down to cloning the source and running docker compose up -d. Google Chrome is the only browser the guide supports. A new account has no rights by default ("You can register a user but by default, it will not have rights even to view the list of tasks"), so the first step after install is creating a superuser from the command line.

The README says CVAT Enterprise is for organizations that need "security controls such as SSO, paid platform features, and SLAs". Enterprise comes as Basic ("Single-instance deployment with essential support for critical issues") and Premium, both by sales contact, with no public price.

Choose it when you want the best documented free route and your team can do without company sign-in. It is the tool most searches for this phrase lead to, with reason.

Label Studio

Label Studio Community is open source under Apache 2.0. It installs with pip install label-studio, with a single Docker command that stores data in a SQLite file, or with Docker Compose and PostgreSQL. It labels many data types, not only images. Users sign up and log in, and an annotation is tied to the account that made it.

Its own edition comparison is direct about what Community leaves out: role-based access control (Admin, Manager, Reviewer, Annotator), assigning reviewers, comments, annotation history, workspaces, SAML SSO or LDAP, and "Activity logs for auditing annotation activity by project". Reviewer assignment and annotation history are in the hosted Starter Cloud plan and in Enterprise. Sign-on and activity logs are in Enterprise, which is available on-premise.

Choose it when you label more than images, or you want the lightest install of a server-based tool. Read the edition table before promising your team review roles on the free edition.

VIA

The VGG Image Annotator from Oxford is "distributed as single HTML files that run offline in a web browser" and does "not require installation or setup". It is released under the BSD 2-clause license. There is no server, so there is nothing to host, and no accounts, roles or review. Shared projects exist in VIA3 through a separate service.

Choose it when one person needs to draw regions on images on a machine that cannot install anything. It is the smallest possible answer to "nothing may leave this computer".

BIIGLE

BIIGLE is "a web service for the efficient and rapid annotation of still images and videos", MIT licensed. A hosted instance runs at biigle.de. To run your own, its admin documentation starts from a clean Ubuntu install: install Docker and Docker Compose, build the production images, start the containers and apply the database migrations. The build needs a GitHub access token, and the setup expects SSL certificates unless it is local.

Choose it when your work is large image collections or video in environmental science, and you have someone comfortable building Docker images.

DataTorch

We make DataTorch, so read this section as the vendor's own.

DataTorch self-hosted is not free and not open source. It is the Enterprise plan. Our Enterprise page says the same product that runs on our hosted cloud "ships as a Helm chart for your own cloud or on-prem, air-gap included", and the pricing page lists "Self-hosted / air-gapped (Helm + offline license)" with "Your storage, your credentials, your compute". A Helm chart means it runs on a Kubernetes cluster, which is a heavier starting point than one Docker Compose file. It is priced per account, and it starts with an email, not a download.

What the product does is the same in either place: annotation in the browser, multi-channel TIFF with a lookup table per channel, review file by file, sign-off of a dataset as a locked version, and an audit trail. Our Enterprise page also says this plainly: DataTorch "is built to run inside your validated environment and to support your qualification work. It is not itself a certified system."

Two things worth knowing before you ask about self-hosting. The Free plan is hosted, and it is the way to try the product on your own images. And if your requirement is that files stay in storage you own, a hosted project can use your own AWS S3, Azure Blob or Google Cloud bucket as its file source (cloud storage links are listed on the Team plan), with no server to run.

Choose it when the team needs accounts, review, sign-off and an audit trail inside its own environment, wants a vendor on the other end, and has a Kubernetes cluster or someone who runs one. If the requirement is free and open source, CVAT Community is the documented route.

A short way to decide

  • One person, one machine, nothing installed: VIA.
  • A small team, free, on a server you already have: CVAT Community, or Label Studio Community if you label more than images.
  • Company sign-in, review and an activity record inside your network: a paid edition. Ask each vendor the questions in the checklist above, in writing.
  • Files must stay in your bucket, and a hosted tool is acceptable: you may not need to self-host at all.

Sources

Read on November 17, 2026.

  • CVAT: the repository README at github.com/cvat-ai/cvat, the installation and upgrade guides at docs.cvat.ai, and cvat.ai/pricing/on-prem
  • Label Studio: the repository README at github.com/HumanSignal/label-studio and the edition comparison at labelstud.io/guide/label_studio_compare
  • VIA: robots.ox.ac.uk/~vgg/software/via
  • BIIGLE: the repository README at github.com/biigle/biigle and the installation page of its admin documentation
  • DataTorch: datatorch.io/enterprise, datatorch.io/pricing and docs.datatorch.io

"Not covered in the pages we read" means those pages did not say. It does not mean the tool lacks the feature.

Try DataTorch on your own images on the hosted Free plan. For a self-hosted or air-gapped deployment, email datatorch@datatorch.io and we reply within a business day.

Frequently asked questions

Is there a free self-hosted image annotation tool?

Yes. CVAT Community (MIT license) and Label Studio Community (Apache 2.0) are free and run on your own machine with Docker. VIA (BSD 2-clause) is a single HTML file that runs offline in a browser. BIIGLE (MIT) can be built and run with Docker Compose.

Do the free self-hosted editions include single sign-on?

Not by their own documentation, read on November 17, 2026. CVAT lists SSO under CVAT Enterprise and its paid online plans. Label Studio lists SAML SSO and LDAP under Enterprise only.

Can DataTorch be self-hosted?

Yes, on the Enterprise plan. The same product ships as a Helm chart for your own cloud or on-prem, with an air-gapped option and an offline license. It is priced per account and starts with a conversation. The Free and Team plans are hosted.

Do I need to self-host to keep images in my own storage?

Not always. Some hosted tools read files from a bucket you own. In DataTorch, a project can use your own AWS S3, Azure Blob or Google Cloud bucket as its file source, and the Team plan lists cloud storage links. If your rule is that no outside service may reach the files at all, you need the tool inside your network.

Try it on your own data

Upload your images and start annotating.

Start free

Related

Tools and community for image annotation: label, review, and share datasets, from everyday photos to the imagery only specialists can read.

Platform

PlatformImaging labsExplore datasetsEnterprisePricing

© 2026 DataTorch. All rights reserved.