Which situation is yours
"Offline" means three different things — here is which one you need
People arrive at this page from three different requirements: a machine with no connection, a rule against uploading, and data that is sensitive enough to need an argument rather than a promise. They have different answers, so they get their own sections.
Is it the best offline image annotation tool? It depends on whether you label alone
If one person — or a few people working on separate things — does the labelling, AnnotateIt is a strong offline choice: install the Windows, macOS or iPad build and the AI engines ship inside it, so the whole loop runs with the network cable pulled out. If several people need to label the same pool of images with review stages, an offline server product such as self-hosted CVAT is the better answer, and this page will not pretend otherwise.
Be careful with the word itself, because that is where most of the confusion lives. A self-hosted server is offline in the sense that data stays inside your network — but a machine still has to run the server; not every product requires containers. A desktop app is offline in the sense that nothing needs to be running except the app. AnnotateIt is the second kind. The browser build is a third case worth stating exactly: it works offline once the app and the engines you use are cached, but the first load needs a connection.
- On a connected machine, install from the Microsoft Store, Mac App Store or App Store — native builds bundle the default engines, so nothing is fetched later
- Carry your existing dataset across as a COCO, YOLO, Pascal VOC or Datumaro archive
- Import, annotate with on-device assistance and export, all without touching a network
- Not the answer when several annotators need one shared queue with assignment and review
No-upload annotation: what actually leaves the machine
The short version: your media, labels and annotations stay on your device, and the AI models run on your hardware. There is no account, no cloud sync and no server-side collaboration — not as a missing feature, but as the architecture. Optional ML jobs are separate: starting one sends its selected dataset to your configured runner.
The precise version, because "no upload" deserves detail: the browser build downloads the application itself, and downloads the AI engines you choose to use, from the network. That traffic goes one way. Your images, videos, labels and annotations are not part of it. Native builds on Windows, macOS and iPad bundle the default engines, so even that download disappears. You can watch this happen — open a network monitor next to the app and annotate; application and model files come down, and nothing carrying your media goes up.
- Local annotation does not upload images or annotations. Optional ChatGPT or Claude and ML runner connections transfer the data described below
- Downloaded: the app itself and any optional engines you choose; native builds bundle the defaults
- Where inference runs: your CPU, or your GPU through WebGPU where available
- Optional AI Assistant can annotate the current image or video frame through OpenAI API or Claude API, or through ChatGPT via Codex or Claude Code on Windows/macOS. Sending a message with the image attached shares its preview, project labels and annotation context with the selected provider (OpenAI or Anthropic). Attachment is on by default in external-assistant mode and can be disabled. Local AI tools remain separate; optional ML jobs send selected dataset bundles to your configured runner.
Private data labelling: what a security review can and cannot lean on
Be clear about what this claims. AnnotateIt is not a certified compliance product, and no annotation tool makes a project compliant on its own — that depends on your controls, your device, your organisation and your regulator. What it removes is the part that usually makes an approval hard: there is no upload, no account and no vendor-side copy of your images, because local annotation needs no hosted backend; optional ML jobs use your configured runner. A long conversation about a new data processor becomes a shorter one about an application running on an already-approved endpoint.
What it does not remove: certification, formal attestations and a DPA covering processing (there is no processing on our side to cover — which answers some reviews and not others), per-user access control, and an audit trail of who saw or changed what. With no accounts there is no user-level log. Device security also matters more, not less, when data never leaves: full-disk encryption, screen lock, backups and physical control of the machine are all yours.
- Fits medical, defence, industrial and personal imagery that cannot leave the network
- Fits pre-publication research data, unfiled IP and air-gapped workstations
- Hand your reviewer the data-flow document and the security overview — both are written for that conversation
- Not sufficient alone when you need certification, access control between users, or an audit trail
