SwiftBox: YOLO Bounding Box Editor

Open an image inside VS Code, drag/resize the green box right on the lesion (or any object), add a second box for multi-object images, swipe through a folder, and save — coordinates are written straight to YOLO .txt, YOLO segmentation, or SwiftBox JSON. No GPU, no server, no upload.

Try it without installing anything

alabed.site/swiftbox runs the same editor in your browser — Chrome, Edge, Brave, Opera, Firefox and Safari. Open a folder of images (or import from Google Drive), annotate, and take the labels away as .txt files, one labels.json, or a zip. The images never leave your machine: everything is decoded and drawn locally, and only labels are ever written anywhere.

Where they go is your choice:

The extension is the same editor with the whole disk available to it, and the two share classes.txt, labels.json and .swiftbox/config.json, so a project moves between them.

Demo

Features

• View any image with its YOLO box drawn in green.
• Drag inside to move, drag a corner to resize.
• Draw bounding boxes (W), polygons (P), or paint masks (B). Tools stay armed, so you can draw shape after shape — V or Esc to go back to select.
• Magic-click segmentation with SAM2/SAM3 (M) — bring your own endpoint. Hidden when no endpoint is configured.
• Change class with the number keys (names read from classes.txt; a digit only works for a class that exists). With nothing selected, the digit arms the class for the next shape you draw.
• Add classes from the editor sidebar — written straight to classes.txt.
• Swipe prev/next image (A / D) — auto-saves before moving.
• Crop (C) and rotate (R) images in-place.
• Undo/redo (Ctrl+Z / Ctrl+Y) with full history.
• Remappable keyboard shortcuts (?).

Output formats

SwiftBox auto-detects the annotation format from your existing project files. In a brand-new project, the editor itself asks you to pick one before you start — no VS Code dialogs:

The choice is saved to .swiftbox/config.json and remembered for the project. Delete that file to be asked again.

Expected layout

dataset/
  classes.txt          # one class name per line
  images/
    img001.jpg
    img002.jpg
  labels/
    img001.txt         # YOLO: "cls cx cy w h" (one line per box)
    img002.txt

classes.txt is found by walking up parent folders (within the workspace), so per-class subfolders work too.

Usage

1. Open the folder of images in VS Code.
2. Right-click an image → Reopen Editor With… → SwiftBox Editor (or set priority to default in package.json to open on click).
3. Edit the box, press Ctrl+S or Enter.

Shortcuts

All shortcuts are remappable via the ? overlay.

License

MIT

Built by Abdulrahman Alabed · alabed.site