Claude and ChatGPT now label for you

From images to training data,on your computer.

Annotation is the slow step between your images and a trained model. The Labelme desktop app drafts each shape with AI, and you correct it.

  • SAM2, SAM3, and YOLO-World draft from a click, box, or word.
  • Labelme uploads nothing, and its AI runs on your computer.
  • It installs without Python and writes plain JSON files.
Try Labelme free
See plans

Available for macOS, Windows, and LinuxTry it free for 3 days without a card. After that you pay once, from $79, and can get a full refund within 14 days.

The Labelme desktop app annotating a synthetic Urban Plaza dataset: a selected dog polygon and the label dialog naming it, shown enlarged in a loupe, with drawing tools, label lists, and dataset files around the canvas.

Labelme annotates any kind of image.

  • AgricultureApples as rectangles
  • MedicalLungs as polygons
  • ElectronicsChip as rectangle
  • PlantsLeaf as polygon
  • IndustrialBearing as circle
  • MicroscopyCells as circles
  • TrafficSign as rectangle
  • ConstructionExcavator as rectangle

Labelme turns your images into annotations you can train on.

Your images
Labelme
The Labelme desktop app annotating a synthetic Urban Plaza dataset: a selected dog polygon and the label dialog naming it, shown enlarged in a loupe, with drawing tools, label lists, and dataset files around the canvas.
AnnotateBy handAI draftReview
Labelme JSON
or convert
Pro toolkitYOLOYOLO-OBBVOC
PyTorchTensorFlow
Train
Your AI model
Works with your training stack
  • PyTorch
  • TensorFlow
  • Ultralytics YOLO
  • PaddlePaddle
  • OpenMMLab
  • Roboflow

Point, box, or words. AI drafts the annotation.

You correct shapes instead of drawing them. Every result stays editable, point by point.

A Labelme parking-lot view with SAM3 oriented boxes around every visible car.

Checked, converted, ready to train.

You check the dataset in the app, then give your training code a format it reads.

Reviewing the annotated orchard in Labelme, zoomed into one selected tree-crown polygon, with a QA flag panel and the full annotation checklist.
ReviewMark an image as checked with a flag, and go through every shape in the annotation list.
Exporting the annotated orchard to Pascal VOC and YOLO with the Labelme toolkit: the terminal commands above the class segmentation masks they produced, tree crowns and truck each in their class color.
ConvertLabelme JSON works as it is. Conversion to YOLO, YOLO-OBB, or Pascal VOC needs Pro and runs in a terminal with Python.

Labelme uploads nothing.

Its built-in models run offline on your computer. Your agent's model may run in the cloud. Connect a local one to keep everything offline.

  • DataLabelme saves images, annotations, and exports in your folders. The app sends no telemetry.
  • Local AIThe models run on your computer, without an API key or a third-party AI account.
  • DownloadsYou download the app with a free labelme.io account, and each model on first use. After that, you can work offline.
Air-gapped or behind a firewall? See the FAQ
Offline
Labelme annotating a synthetic medical image on the desktop.

Open a folder and start annotating.

Everything in one window.

There is nothing to configure before the first image.

The Labelme window: tool column, canvas, label list, and file list.

7shape tools.

Pick a tool with one click, then draw.

  • Polygon
  • Rectangle
  • Oriented rectangle
  • Circle
  • Point
  • Line
  • Line strip

One key to the next image.

  • DNext
  • ABack
  • THide shapes

Saves as you work.

There is no save step.

  • On every edit
  • On image switch

21languages.

English简体中文繁體中文日本語한국어DeutschFrançaisEspañolPortuguêsItalianoNederlandsPolskiРусскийУкраїнськаTürkçeΕλληνικάMagyarBahasa IndonesiaTiếng Việtไทยفارسی

Flags and notes.

Mark an annotation occluded or truncated, or add a note.

AI when you want it.

Or draw every shape by hand.

A Labelme parking-lot view with SAM3 oriented boxes around every visible car.

Plain JSON files.

Labelme writes one beside each image, and any script can read it.

{"imagePath":
   "urban_plaza_0001.png",
 "shapes": [{"label": "dog",
   "shape_type": "polygon",
   "points": [[832.4, 861.0], …]}]}

Your AI agent works right in Labelme

  • ChatGPT outlines every field in the aerial photo

    create_shapes

  • Claude selects the boxes that miss their car

    select_shapes

  • Claude Code deletes the duplicate labels

    delete_shapes

  • ChatGPT looks at the image you have open

    get_image

  • Claude reads every shape and label

    get_state

Connect ChatGPT, Claude, Claude Code, Codex or Cursor from File › Connect AI Agent… in Labelme.

Set up your agentSee it work

Already use open-source Labelme? Here's what Pro adds.

The open-source package has the same annotation tools and AI models. Pro adds AI agents and YOLO and VOC export, and installs without Python.

Open-source packageLabelme app
AI agentsNo agent connectionOne switch lets ChatGPT, Claude, or Claude Code work on the image you have open
Dataset exportLabelme JSON. Converting it to YOLO or VOC is up to youThe dataset toolkit exports to YOLO, YOLO-OBB, and Pascal VOC
InstallSet up Python, then run pip install labelmeOne installer for macOS, Windows, or Linux. Double-click it to open the app. It needs no Python.
Model downloadsWeights download from GitHub and Hugging Face, which filtered networks can blockWeights download from the Labelme mirror, so a filtered network needs to allow one host
Updates & supportYou upgrade with pip and resolve package conflicts yourselfYou get a year of tested builds, and your version keeps working after that. Email support is included.
CostThere is no license fee. Setup and upkeep cost your time.One payment from $79

"I used the free version before, and saved a lot of hassle, now, this wouldn't be much to invest."

Yoon Young JoonVerified purchase
Try Labelme freeSee plans

Rated 4.9 from 146 reviews.

Every review is from a buyer, quoted as written.Read them all on Gumroad

5 out of 5 stars

"The best thing about it is that it is easy to annotate, even for the average person who is not a programmer. Other apps require writing code or complicated steps, but this has none of that."

WakuiMoritoshi
5 out of 5 stars

"Very awesome tool, helped me create lots of datasets by myself in a very efficient and quick way! Easy to use and performs outstandingly, very happy"

John Smith
5 out of 5 stars

"The best standalone annotation tool."

Royi Avital
5 out of 5 stars

"Super useful for image annotation in Windows."

Gumroad review

From the person who wrote Labelme.

A crowd watching a red two-armed robot pick items at the Amazon Picking Challenge.
The Amazon Picking Challenge, where Labelme began.

I wrote the first version of Labelme in 2016, as a student in a robotics lab at the University of Tokyo. Our team was entering the Amazon Picking Challenge, and I needed annotated images to train the robot's vision. It has been open source ever since, and I still maintain it.

The app on this page is that same Labelme, built and tested so you can skip the setup, with AI agents and dataset export added. Buying it pays for that work and keeps the open-source project maintained. If it does not work for you, email support@labelme.io within 14 days and I will refund you in full.

Kentaro WadaAuthor of Labelmekentaro@labelme.ioPhD from Imperial College London. Formerly computer vision lead at Mujin. Alumnus of the JSK Robotics Lab, University of Tokyo.

Frequently asked questions

What's the difference between open source and Pro?

Pro adds an MCP server, so any AI agent can annotate for you. It also adds a model download mirror that stays fast behind firewalls, YOLO and Pascal VOC export, priority email support, and a desktop app with no Python setup. The annotation tools and AI models are the same. Compare the plans.

Can I try Labelme before buying?

Yes. Download the app and use it free for three days, with no credit card. The trial excludes the Pro dataset toolkit.

Is Labelme a subscription?

No. You pay once and keep your version forever. Pro includes a year of updates, and Pro Lifetime includes every future update.

Do my images ever leave my computer?

Not through Labelme. The app uploads nothing and sends no telemetry, and its built-in models run offline on your computer. Your agent's model may run in the cloud. Connect a local one to keep everything offline.

What is the refund policy?

Email support@labelme.io within 14 days of payment for a full refund, no questions asked.

Can I run Labelme fully air-gapped?

Yes. Download the AI models on a connected computer first, or ask support@labelme.io for the offline pack.

Which AI models are supported?

SAM, SAM2, SAM3, and EfficientSAM for segmentation, plus YOLO-World for text-prompt detection. Every model runs on your computer.

Which export formats are available?

Pro exports to YOLO, YOLO-OBB, and Pascal VOC, and imports YOLO and YOLO-OBB datasets back.

Annotate your next dataset on your computer.

AI drafts each shape on your computer, Labelme uploads nothing, and one installer replaces the Python setup.

Try Labelme free
See plans

Available for macOS, Windows, and LinuxTry it free for 3 days without a card. After that you pay once, from $79, and can get a full refund within 14 days.