Lannotex

Label images for AI, on your own machine.
Boxes, polygons and classes. Any image size. No account, no telemetry, MIT.

Documentation · Quick start · Coming from LabelMe · Troubleshooting · Releases

Why Lannotex?

• It stays quick on dense images. Editing costs the shape you touched, not the size of the document. An image with 200 shapes used to freeze for six seconds on every mouse release; the same edit now takes 0.2 ms. CI fails the build if an edit ever exceeds 16 ms.
• Image size stops mattering. Zooming past the preview builds a tile pyramid once and caches it. A 900 MP TIFF reopens in 0.1 s and pans using under 200 MB — decoding it whole would need 2.7 GB. (measurements)
• You can't lose work. Every edit is written to a recovery draft within 250 ms, saves are transactional, and undo history survives a crash. Self-intersecting polygons and zero-area boxes are refused before they reach your dataset.
• Everything you need is free. Every export format, the desktop builds and team review are all MIT. No licence keys, no Pro tier.
• Private by default. No account, no telemetry, no cloud. It makes network requests only if you choose to host or join a team on your own network.

Install

uv tool install lannotex   # recommended, isolated install
lannotex

pipx install lannotex
# or, inside a virtual environment
python -m pip install lannotex

git clone https://github.com/Lourdhu02/Lannotex.git && cd Lannotex
uv venv && uv pip install -e ".[dev]"
lannotex

Lannotex supports Python 3.11–3.14 on Windows, macOS, and Linux. See the installation guide for details.

Your first dataset in 60 seconds

1. Create a project and pick a pipeline: Classification, Object detection, Segmentation, or Combined. Each one shows only the tools it needs.
2. Import a folder (Ctrl+I). JPEG, PNG, WebP, TIFF and BMP files are checked and referenced in place, not copied.
3. Add labels, then draw: B for boxes, P for polygons (Enter finishes), V to select and edit.
4. Save & next with Ctrl+Enter. Autosave and recovery drafts run in the background anyway.
5. Export to the format your training code expects.

Every shortcut can be remapped. See the annotating guide for the full list.

Light theme · polygon vertex editing	Dark theme · object detection

Features

Lannotex vs. LabelMe

LabelMe is a great tool and Lannotex reads and writes its JSON format, so you can switch without converting anything.

Based on LabelMe's published features and pricing as of October 2026. The full comparison and migration guide covers which LabelMe shapes import today.

Command line

lannotex new ./dataset                                  # create a project
lannotex analyze ./dataset                              # JSON health report: counts, problems, drafts
lannotex export ./dataset ./out --format "YOLO boxes" --split   # also: YOLO segments, Pascal VOC, COCO JSON, Labelme JSON, Classification CSV, Native JSONL
lannotex invite ./dataset "Alice"                       # print a one-time reviewer token
lannotex host ./dataset --bind 0.0.0.0 --port 8765      # share a project on your LAN (TLS via --cert/--key)

analyze writes clean JSON to stdout, so it can gate a training pipeline. See the CLI reference.

Privacy

Solo projects make no network requests. Team mode connects only to the host address you enter, reviewer tokens are kept in memory on clients, and the host stores only token hashes. There is no telemetry, no external AI service, and no automatic upload. Read the team guide for LAN and TLS setup.

Roadmap

Next up: point, line, rotated-box, keypoint and brush-mask tools, per-object attributes, and VOC, YOLO-seg and YOLO-OBB export. See the roadmap and the measured performance.

Contributing

Bug reports, ideas, and pull requests are welcome. Start with CONTRIBUTING.md, and please follow the Code of Conduct. Report security issues privately, as described in SECURITY.md.

If Lannotex saves you time, a ⭐ helps other people find it.

License

MIT © 2026 Lourdhu Raju. Third-party licenses are listed in THIRD_PARTY_NOTICES.md. Demo images are public-domain and CC0 samples from scikit-image.