AnnoLabel

Local CLI · by Andesprit

Label images.
Review the result.
Export COCO.

Bring your vision-capable agent. AnnoLabel gives it the tools to save image classifications, bounding boxes and segmentation polygons.

Install AnnoLabel Read the guide

Python 3.11–3.14 · macOS, Linux and Windows · MIT code

Concept illustration of bounding boxes and polygons being reviewed before COCO export.
Concept illustration. AnnoLabel is a command-line tool; the agent supplies the labels and geometry.

One image, one reviewable task

From an image to a dataset.

  1. 01

    Prepare

    Create a task with your image, labeling instructions and optional object categories.

  2. 02

    Submit

    Your agent opens the image and supplies a complete annotation snapshot. AnnoLabel checks the geometry and stores it.

  3. 03

    Review

    Open the annotated image. Check omissions, labels and outlines; use the next task handle for a correction when needed.

  4. 04

    Export

    Export COCO instances and oriented images, with scene classifications and provenance in separate files.

Get started

Install once.
Work locally.

With uv installed, get the CLI from PyPI. Your agent needs image viewing and shell access.

The CLI runs locally and makes no model calls. Any model service used by your agent depends on that agent’s setup.

Terminal

uv tool install annolabel
annolabel --help

Command not found? Run uv tool update-shell and restart your shell.

View the package on PyPI
AnnoLabel's synthetic example showing a rectangle and a triangle with annotation outlines.
Actual AnnoLabel output from the project’s reproducible shapes example.

Try the included example

See the output before using an agent.

The repository includes an image, task instructions, categories and a complete submission. Run the example to inspect the review image and exported dataset.

Run the shapes example

Know the limits

Simple polygons are supported. Holes, multipart instances, video and keypoints are outside the current scope. Geometry validation does not establish visual accuracy; review remains necessary.