Overview
Labelme is a graphical image annotation tool that stores each annotation as a JSON file beside its image.
According to the maker's site, Kentaro Wada wrote the first version in 2016 as a student at the University of Tokyo and still maintains it. The privacy policy and the license agreement name Kentaro Wada, trading as Labelme.io, as the operator. The repository describes the project as image annotation with Python, supporting polygon, rectangle, circle, line and point shapes. Its README also lists image flags, a video-annotation example, AI-assisted point-to-polygon annotation with SAM and EfficientSAM, and text-to-annotation with YOLO-World and SAM3.
Version 7.11.0 was released on 14 November 2026 and is the latest GitHub release. Version 7 of the pip package needs Python 3.12 or newer and uses PySide6 (Qt6); a v6.3.x line receives critical fixes only, per the README.
The maker positions the standalone app as the same Labelme, packaged so you skip the Python setup, with AI-agent access and dataset export added. These are maker statements. PromptQuorum did not run the app and did not measure annotation speed or model quality.
How to Get It
The easiest route is the standalone app, a normal installer for macOS, Windows or Linux; a pip install of the open-source package also exists.
The download page lists a .dmg for Apple Silicon and Intel Macs, a Microsoft Store listing, an .msi installer and a portable EXE for Windows x64, and an AppImage for Linux x64. The installation guide marks downloads as sign-in required. Windows may show a "Windows protected your PC" warning; the maker says to confirm that the publisher includes "Kentaro Wada" and not to disable Windows Security. The GitHub repository lists no release assets (checked 18 November 2026), so the installers come from labelme.io.
Platform | How to get it |
|---|---|
| macOS | Download .dmg (Apple Silicon or Intel), move to Applications |
| Windows | Microsoft Store, .msi installer or portable EXE |
| Linux | AppImage: make it executable, then run it |
| Python | Run pip install labelme in a terminal (Python 3.12+) |
| Website | labelme.io |
| Privacy policy | labelme.io/privacy-policy |
The pip route needs a terminal and a Python 3.12 or newer environment; the standalone app needs neither.
Requirements
The maker does not state RAM, GPU or disk requirements for the app.
The installation guide says Labelme supports operating system versions that are not end-of-life. AI model weights download on first use, and no download sizes are given. An offline SAM3 pack exists for machines without internet access, per the same guide.
Item | Stated figure |
|---|---|
| Operating system | macOS (Apple Silicon, Intel), Windows 64-bit, Linux 64-bit |
| RAM | Not stated |
| GPU | Not stated |
| Disk and model size | Not stated; weights download on first use |
| Python (pip route) | 3.12 to 3.14 for v7.x (README) |
| Network | Needed for sign-in download and model downloads |
The license check itself runs offline, according to the maker.
Features
Labelme combines manual shape tools with AI drafting, and keeps every result editable. All items below are maker claims from the website, changelog and README.
- Seven shape tools: polygon, rectangle, oriented rectangle, circle, point, line and line strip.
- AI drafting: click-to-polygon with SAM2, box-to-annotation and text-to-annotation with SAM3, and YOLO-World for text-prompt detection.
- Models named by the maker: SAM, SAM2, SAM3 and EfficientSAM for segmentation, plus YOLO-World.
- Flags for occluded or truncated shapes, notes, shape locking and a flag to mark an image as checked.
- Auto-saving to a JSON file beside each image, and keyboard shortcuts to move between images.
- An interface in 21 languages.
- Pro only: an MCP server so ChatGPT, Claude, Claude Code, Codex or Cursor can read, add, select and delete shapes on the open image. You approve each change, and one undo reverts it.
- Pro only: a command-line dataset toolkit with 10+ tools for batch annotation, review and export to YOLO, YOLO-OBB and Pascal VOC. It runs in a terminal and needs Python.
Privacy and Data Flow
Per its privacy policy, the app uploads none of your images or annotations, but several features use the network.
The policy, last updated 12 November 2026, says the app checks the license key offline and does not check for updates. It does go online to download the AI models you choose. The files come from a Cloudflare-hosted Labelme server or, as a fallback, from GitHub or Hugging Face, which see your IP address.
If you connect an AI agent, the policy says the agent receives the open image, its file path, the annotation file path, the labels, and each shape's label, points, flags and description. The agent's provider governs that data, while an agent that runs locally keeps it on your computer.
Application telemetry is opt-in. It covers app opens, trial status, Buy clicks, license activation and unexpected errors, goes to PostHog under a random ID, and can be switched off from the app menu. The website uses Google Analytics and PostHog, and sign-up stores referral details. Paddle processes payments as merchant of record. The operator is based in Japan, and its providers process data in other countries, including the United States.
Pricing and License
The open-source package is free under GPL-3.0, and the standalone app is a one-time purchase under the maker's license agreement.
- Open-source package: free, installed with pip. The maker says it has the same annotation tools and AI models as the app.
- Pro: $79 once, with one year of updates; the version you have keeps working after that. It adds the AI-agent connection, the dataset toolkit, a model-download mirror and email support, per the maker.
- Pro (Lifetime): $249 once, with every future update, major releases included.
- Team: $249 per seat for five or more seats that can be reassigned; quotes on request.
- Trial and refund: 3 days with no credit card, excluding the dataset toolkit; a full refund by email within 14 days of payment.
Who Should Use It
Labelme fits people who annotate images for computer-vision models and want to do it on their own computer.
- Machine-learning practitioners who want images kept off third-party annotation servers.
- Annotators who prefer correcting AI-drafted outlines to drawing every shape by hand.
- Users of ChatGPT, Claude or Claude Code who want an agent to edit shapes and accept that agent's provider handling the data.
- Python users who are happy with the free package and a terminal.
What We Could Not Verify
Labelme is a poor fit in some situations, and several details remain unverified.
- Not a fit: teams that need a shared web annotation server with task queues. The sources describe a single-user desktop tool and no collaboration server.
- Not a fit: air-gapped setups that cannot fetch the model weights or the offline pack first.
- Unverified: RAM, GPU and disk requirements and model download sizes. None are stated.
- Unverified: annotation speed and AI model accuracy. PromptQuorum measured neither.
- Unverified: which export formats the app offers without Pro. The README lists VOC and COCO export among the open-source features, while the maker's site says YOLO, YOLO-OBB and Pascal VOC export belong to the Pro toolkit.
- Unverified: the homepage figures of 920+ paid licenses and a 4.9 rating from 146 reviews are the maker's own statements and were not checked.
- Unverified: that the agent connection works with every listed agent, and how the Linux AppImage behaves on specific distributions.
Labelme vs Alternatives
Labelme is a single-user desktop annotator. PromptQuorum has not compared it hands-on with the alternatives below. Their details come from their GitHub descriptions, checked 18 November 2026.
Labelme
- Noted in sources:
- Desktop app; GPL-3.0 package plus paid app
CVAT
- Noted in sources:
- MIT; image, video, audio and 3D annotation platform
Label Studio
- Noted in sources:
- Apache-2.0; multi-type data labeling tool
X-AnyLabeling
- Noted in sources:
- GPL-3.0; cross-platform desktop annotation app
FAQ
What happens when the year of updates ends?
Nothing stops working. You keep every release from your year of updates, and the app checks the license key offline. You can buy another year or choose the lifetime plan, per the maker.
Do my images leave my computer?
Not through the app itself, according to the privacy policy. Model weights are downloaded, and an AI agent you connect may send the data to its provider unless it runs locally.
Can I use Labelme without an internet connection?
After the models are downloaded, the maker says it works offline. For air-gapped machines, it says to download the models on a connected computer first or to ask support for the offline pack.
Which AI agents can connect?
The maker lists ChatGPT, Claude, Claude Code, Codex and Cursor, connected from File, then Connect AI Agent, in the Pro app. The changelog says you approve each change.
Is the standalone app open source?
No. The source code and the PyPI package are GPL-3.0, but the app is distributed under the maker's license agreement, which forbids redistribution and reverse engineering.
Verdict
Labelme is worth trying if you annotate images for computer-vision models and want the work and the AI models on your own machine.
It has a long track record, with a first version in 2016 and a repository commit on 17 November 2026. It also offers a free open-source route, offline AI drafting by the maker's account, and plain JSON output. The paid app removes the Python setup.
The trade-offs are real. It is one user on one desktop, with no collaboration features described. Hardware needs and speed are unverified, the download requires a sign-in, and the agent features can pass image data to a cloud provider. PromptQuorum has not tested the app, so try the trial or the pip package on a few images before you buy.
Sources
- Labelme website — maker's description of the app, AI models, agents and offline use, checked 18 November 2026.
- Pricing page — plans, prices, trial, refund and license terms, checked 18 November 2026.
- Download page — per-OS downloads, version 7.11.0 and Windows notice, checked 18 November 2026.
- Installation guide — install steps, sign-in, trial, model downloads and offline pack, checked 18 November 2026.
- Changelog — Pro AI-agent feature and shape tools, checked 18 November 2026.
- Privacy policy — data flow, telemetry, agents, operator based in Japan, checked 18 November 2026.
- License agreement — license scope, restrictions, model licenses and governing law, checked 18 November 2026.
- About page — maker background and open-source history, checked 18 November 2026.
- Labelme GitHub repository — project description, GPL-3.0 license, last commit and README, checked 18 November 2026.
- Release v7.11.0 — latest release version and date, checked 18 November 2026.
- CVAT repository, Label Studio repository and X-AnyLabeling repository — licenses and descriptions of the alternatives, checked 18 November 2026.
