Private · runs on your device
Collect images and video, annotate with local AI, ChatGPT or Claude, and turn reviewed datasets into measured results. Local-first storage, with an optional online chat for annotation.
Workflow
Collect, curate, annotate, review and evaluate locally. Export a reproducible dataset or send a selected version to your own ML runner.
Files, camera, screen or a recording
Sources →One click, a text prompt or a pose box
Smart tools →D-FINE · RT-DETR · DEIM · RF-DETR over the whole set
Models →Quality scan, train / val / test, freeze
Dataset ops →Test-set metrics, exports or an optional runner
Evaluate →Pick a task: detection, segmentation, keypoints or classification.
Import files or datasets, or collect from a camera, screen or video-file source. Inspect staged captures before accepting them.
Select useful frames, filter by source, crop, resize, adjust colour or redact image regions before labelling.
Label manually or let on-device AI do the heavy lifting.
Accept or reject AI drafts, run a quality scan, assign train/validation/test, then save a version you can diff or restore.
Measure an available model on reviewed test images. Export the dataset, or use the experimental desktop ML runner workflow.
Annotate with ChatGPT or Claude
Ask for boxes, segmentation polygons or image classes in plain language. Work with one image or video frame in chat, or give Auto-annotate one instruction for a batch of images. Review the results directly in AnnotateIt.
Example request“Create a Dinosaur label, then outline every dinosaur in this image.”
Open Annotate with ChatGPT or Claude above Next. The current image or frame is attached by default; you can turn it off.
Create, rename, recolor or delete project labels with confirmation. Ask follow-up questions to refine pending boxes or polygons.
Review the usual AI suggestions on the canvas. Use the existing Accept / Reject controls; the chat reports when nothing was annotated.
Optional online feature: an attached chat message sends one image or one frame and context to your selected provider, OpenAI or Anthropic. Starting an assistant batch sends eligible images and project labels from your chosen scope. No whole-video or keypoint generation. Local AI tools remain separate.
A computer vision platform
Annotation is one stage. Connect a source, select useful frames, turn AI suggestions into reviewed labels, preserve the dataset and measure the result.
Camera, screen, video-file and, on desktop, RTSP/HTTP network-camera sources feed one capture workflow. Use timelapse, motion or Smart selection; inspect the staged results before accepting them.
Explore capture sources →Trace captured media to its source and session, see why a frame was kept, and filter by source or capture reason before annotating or exporting.
Trace your data →Evaluate an available prediction engine against reviewed test images. Inspect task-specific metrics, per-label results and the exact data behind the score.
Evaluate models locally →Export standard datasets or use the experimental desktop ML runner integration for a versioned RT-DETR fine-tuning and inference workflow.
Understand ML Pipelines →Capture availability depends on the platform. Network cameras need the desktop app; watched folders are not implemented. Optional ML jobs send the selected dataset to your configured runner.
Local-first
AnnotateIt AI stores projects, media, labels and annotations on your device. Local AI models run on your hardware; optional external AI connections send the content needed for your requests to the selected provider (OpenAI or Anthropic).
Images and videos you import stay in local storage on your device.
Manual tools and on-device AI models annotate locally. No account, no cloud backend.
You export standard dataset formats to wherever you choose.
Who it’s for
AnnotateIt is a single-user tool by design — no accounts, shared team queue or multi-user approval workflow. AI drafts still have a local review queue before they become ground truth.
Collect, annotate and evaluate data without an annotation server or account. Export to your own training stack, or prepare a separate ML runner for the experimental desktop fine-tuning workflow.
How it comparesPre-publication images and unfiled IP stay on the machine that produced them. Dataset versions, seeded splits and quality checks all run locally, so the dataset behind a paper stays reproducible without a hosted platform holding it.
Private data labellingField laptops, restricted networks, client material under NDA. The native builds bundle the default AI engines, so the whole loop — import, AI-assisted labelling, export — runs with no network at all.
Security overviewNot your situation? If several people label from one shared queue with assignment and review stages, a server-based tool is the better answer — and the comparison says which one. See the comparison
Annotation tasks
Create a project for the task you are training for — every project type has purpose-built annotation and video workflows.
New to computer vision?What image annotation is, and which type fits your dataset →
Draw axis-aligned boxes, refine model suggestions, and keep object identity with video tracks.
Create object-by-object masks and polygons, then keep their identity with video tracks.
Skeleton templates with joints and edges, with pose assistance for people.
Single-label, multi-label and hierarchical projects, with frame ranges for video.
Import images including HEIC/HEIF and TIFF as stills, or supported CFR or VFR video. Capture photos or silent clips, then annotate with tracks, frames or ranges.
On-device AI
Every model below executes on your machine. Built-in engines are ready immediately; larger ones download on demand — and still run locally.
One-click segmentation
Click an object to get an editable polygon by default, refine with clicks, or choose Mask to keep the model’s exact pixels. MobileSAM is built in; SAM 2.1 Tiny, Small and Large and SAM 3 Tracker are optional where the platform supports them.
Text-prompt annotation
Describe each label once. The active MobileSAM, SAM 2.1 or SAM 3 engine proposes regions and CLIP or SigLIP 2 ranks their similarity; in detection projects Grounding DINO can instead predict boxes directly. The toolbar names the exact pair that will run.
Semantic search & batch pre-labelling
Describe what you need — "a red car at night" — and every matching image in the dataset is ranked, with a heatmap of where it matched. Turn the result set into a batch run, or launch Auto-annotate from a dataset, selection or filter.
Classify without training, using an engine you never had to fine-tune. CLIP is built in; SigLIP 2 is the stronger, multilingual upgrade.
Draw a person box and get 17 COCO keypoints placed onto your skeleton template — several people per image, each its own skeleton.
Box one example and every region of the same image that looks like it comes back as a proposal. A similarity routine, not a model — nothing to download, and it works on a picture full of repeats.
A prediction source rather than a tool: it takes the objects you have already annotated as references and looks for similar ones, running on whichever segmentation engine is active.
Ready-to-use object detectors and instance segmenters, all trained on the 80 COCO classes. Download one from the Models page and set it up in a project — match labels, test on one of your own images, save — with the checkpoint verified on your device first. Every one runs locally on your CPU.
Import your own model through a guided wizard — inspect, map labels, test, save. Supports YOLOv8 detection, segmentation and pose, RT-DETR, DETR, RF-DETR-style and EdgeCrafter-style segmenters, and image classifiers; unsupported output shapes are rejected. Experimental.
Engine availability depends on platform and hardware. CPU/WASM is the baseline, while SAM 2.1 Large, SAM 3 Tracker and Grounding DINO require the app’s WebGPU capability. Mobile omits downloadable engines. Windows, macOS and compatible desktop browsers offer those three engines with a working WebGPU adapter. Models marks incompatible choices as unavailable and still shows their licence, size, speed and accuracy.
Model walkthrough
Segmentation, detection, tracking, poses & search
Real product
Every screenshot below is the shipping application.
Dataset management
Versioning, quality checks, review queues, combined exports, splits and an audit trail — running against a local database on your own machine, with no infrastructure and no tier gate.
Score an available engine on reviewed test images. Inspect task-specific and per-label metrics, with the exact benchmark scope saved alongside each result.
Keep source, session, capture reason and available recording timestamps with collected images. Filter the dataset by how its media was collected.
Freeze a dataset — media, annotations, tracks, labels, attributes and the split — as an immutable snapshot. Compare two versions, download one as a dataset, or restore it. Free on every tier.
Scan for unannotated media, broken or out-of-bounds shapes, label problems, missing required attributes, exact duplicates, odd image dimensions, unreadable media, suspiciously tiny objects, class imbalance and thin video coverage — before you export.
Deterministic, seeded, optionally stratified splits with manual overrides and locks. Exports lay out the native split structure per format and carry a manifest so the exact split can be restored.
Filter by source, capture reason, name, type, size, dimensions or date — or search semantically for image content, with a match heatmap.
An append-only record of what happened and when: media imported, edited or deleted, annotation batches, label changes, dataset copies, splits applied, versions created and restored.
Crop, resize, rotate, flip, adjust colour, and redact image regions with blur, pixelation or a solid fill — opened straight from the dataset.
Duplicate a dataset with or without its annotations — a cancellable operation that rolls back cleanly and records itself in the history.
A batch shows the current image, actual inference stage and elapsed times while it runs, then reports duration, average inference time, saved annotations, empty results and failures. Drafts stay isolated from ground truth until you review them in the gallery or annotator.
Export labels as a small AnnotateIt JSON file — colours, shortcuts, prompts, attributes and hierarchy included — then import them into a compatible project. Match existing labels to preserve their IDs and annotations, or create new ones without overwriting the current schema.
Merge selected image datasets into one archive, preserving their splits, recombining them safely or writing a flat dataset. Byte-identical images never leak across train and test.
Change a project type only after a calculated preview shows the regions created, annotations dropped and labels re-scoped — then export the original first if the conversion is lossy.
Export one project as a portable archive, or back up your entire local profile — every project, all media and the models you downloaded — to a single file you keep.
Local-first means the device holds the only copy. Versions and history are a working record, not a backup — the archive is what protects you from losing the machine. How multiple datasets and combined export work
Interoperability
Move annotations between AnnotateIt and your training stack with standard dataset formats.
Availability depends on project type and dataset contents — Supervisely Video, MOT and MOTS, for example, apply only when a dataset contains video, and KITTI is offered for image detection.
Dataset format details →Supported image and video files →Platforms
Start in the browser in seconds, or install a native build for offline work and store-managed updates.
Runs in a modern browser. Models download on demand and are cached locally.
Open web appNative desktop build with bundled default engines, CPU/fallback downloadable models and the local REST API. WebGPU-gated models are available when WebView2 exposes a working adapter.
Microsoft StoreNative build for Apple silicon Macs with the full annotation workflow, local REST API, custom ONNX import and semantic search. WebGPU-gated engines are available when WKWebView exposes a working adapter.
Mac App StoreTouch-first annotation on the go — free, with no project limit and nothing to buy.
App StoreYes. Capture supports cameras, screen sharing where the platform exposes it, playable local video files and, in the desktop app, RTSP/HTTP network cameras. Choose manual, timelapse, motion or available Smart rules and inspect the staged results before Accept. Watched-folder ingestion is not implemented.
Evaluate an available prediction engine locally on reviewed test images, with task-specific metrics and saved benchmark scope. Fine-tuning is a separate experimental desktop integration with a prepared TAO RT-DETR runner. ML jobs send the selected dataset to that runner; the hosted web app cannot connect through the current transport policy.
Local annotation stores datasets on your device and runs smart tools there. Optional AI Assistant sends chat context and the attached image/frame, labels and annotation context to the selected provider (OpenAI or Anthropic) through your chosen connection. Starting an optional ML job sends the selected dataset bundle to your configured runner, which may be on your machine or remote.
No. There is nothing to sign up for — open the app and create a project.
Once the app and the models you use are cached, annotation work runs locally. The native desktop builds are the best choice for fully offline sessions.
The same product. Native builds bundle the default AI engines, work offline and get store-managed updates; the web version downloads engines on demand and is the fastest way to try AnnotateIt.
Object detection, instance segmentation, keypoint detection, and single-label, multi-label and hierarchical classification — for images and video frames.
Bounding boxes, circles, polygons, open polylines, pixel masks and skeletons, depending on the project. The rotated-box tool requires an existing rotated-detection project; the current Create project UI does not offer that subtype. Polygons and polylines rotate with a handle. Segment Anything creates an editable polygon by default; choose Mask output for exact pixels.
JPG/JPEG/JFIF, PNG, BMP, WebP, TIF/TIFF, HEIC/HEIF, AVIF and GIF. TIFF uses its first page, HEIF its first still, and GIF/AVIF their first frame; AVIF needs a compatible browser or WebView decoder. Previews are 8-bit, while untouched original files remain available for download and export.
CFR or VFR MP4, MOV, M4V and WebM with readable timing and supported platform codecs. AVI/MKV frame extraction requires the native desktop decoder; playback can be unavailable. Variable-frame-rate video is supported with average FPS and exact frame indices; imported bytes and timing stay unchanged.
No. Untouched original image bytes are preserved in downloads, dataset exports and project backups; export does not replace the stored originals. Archive filenames can follow format or collision rules. An explicit image edit/save is different: TIFF, HEIC/HEIF, AVIF, GIF and BMP become an edited PNG with a matching extension and MIME type.
Export supports COCO, YOLO, Pascal VOC, Datumaro, Supervisely Video, MOT Challenge, KITTI, MOTS and Plain ZIP (media only), depending on project type and whether the dataset contains video; import auto-detects COCO, YOLO, Pascal VOC and Datumaro archives.
Yes, and all of it is free on every tier. Dataset versions are immutable snapshots you can diff, download and restore; the Quality tab scans for unannotated media, broken shapes, duplicates, class imbalance and more; deterministic train/validation/test splits export in each format’s native layout; and an append-only activity history records what happened to the dataset.
Yes. Semantic search indexes your dataset on your device once, then ranks images against a plain-language query and shows a heatmap of where the match came from. The same query can be turned into a batch run that pre-labels every match for review.
AI Assistant generates editable boxes, segmentation polygons and image classifications, and can create, update or delete project labels with confirmation. Review annotations with the usual Accept / Reject controls. Use OpenAI API or Claude API on web/desktop, or ChatGPT via Codex or Claude Code on Windows/macOS. It is unavailable on native iOS/Android. Sending a message with the current image attached shares its preview, labels and annotation context with the selected provider (OpenAI or Anthropic); no whole video is sent.
Yes. Alongside the built-in tools, a curated set of auto-annotation models — the D-FINE, RT-DETR, RT-DETRv2, DEIM and RF-DETR detectors plus EdgeCrafter ECDet, with EdgeCrafter ECSeg and RF-DETR Seg for instance segmentation — downloads from the Models page and sets up inside a project: map its classes to your labels, test on one of your images, save. All are trained on the 80 COCO classes, verified on your device before they save, and run locally on your CPU. No weights to supply and no training — that is what separates them from importing your own model.
Yes — an experimental five-step wizard, opened inside a project, imports YOLOv8 detection, segmentation and pose models, RT-DETR, DETR and plain image classifiers (up to 512 MB). It inspects the file, maps its classes to your labels and requires a successful test run before saving. Other output shapes are rejected — it is not "any ONNX".
The experimental desktop Pipelines integration can submit a frozen image-detection dataset to a separate TAO RT-DETR runner for fine-tuning and inference. It needs a prepared runner environment. The hosted web app does not train models; standard exports work with your own training stack.
Yes. Unlimited projects are included in Free on web, Windows, macOS, iPhone and iPad, together with the annotation workflow, the AI tools supported by that device and import/export.
Built by an experienced computer vision engineer
AnnotateIt AI was created by Yuriy Volokitin, drawing on more than 15 years of software engineering experience and hands-on work with computer vision, annotation and dataset-management products, including engineering leadership associated with Intel Geti.
AnnotateIt is designed around the workflow its creator wanted for practical dataset work: local processing, direct control of media and annotations, useful AI assistance, established export formats, and no mandatory account or cloud service.
Actively developed
Latest documented web release6.16.12
Releases are written up here on the site — what changed, in which version, on what date. What is planned next, and what has been ruled out on purpose, is public too.
Create your first local project in the browser. Optional ChatGPT or Claude annotation needs your own AI connection.
Questions before you start?Contact support →
