Getting Started with Cowart on Codex: Full Installation, Canvas Basics, and Image Annotation Editing

When first encountering Cowart, the plugin is not listed in Codex's official Plugin Directory because it must be installed from its own Git marketplace. The article explains how to register the marketplace and add the plugin using the commands codex plugin marketplace add zhongerxin/Cowart --ref main and codex plugin add cowart@cowart-github, then verify activation with codex plugin list.

After installation, a new Codex task loads the Cowart skills and MCP tools. Opening the Cowart canvas is done by invoking the cowart-open-canvas skill, which displays an infinite drawing area backed by the canvas/ folder of the current project. The folder structure looks like:

canvas/
└── pages/
    └── <page-id>/
        ├── cowart-canvas.json
        └── assets/
            ├── image.png
            └── page.html

The canvas UI is divided into several regions: a central drawing area, a top management bar (page switching, undo/redo, copy, delete), a bottom toolbar (selection, pan, drawing, text, annotation, and shortcuts for AI Image, AI HTML, AI Slides), a right‑hand style panel, and a lower‑left view panel. The AI generation panel accepts a prompt, reference images, and size settings, and its quick‑action buttons adapt to the currently selected object.

To generate an AI image, click the AI 图片 button on the bottom toolbar, enter a descriptive prompt, and Codex receives the prompt together with the image‑frame dimensions. Codex then creates the image, replaces the AI frame with a regular image shape, and places the result beside the original for easy comparison.

For image editing, add arrows and short text annotations on the original picture, then click the 按标注修改 button. Cowart captures a screenshot of the annotated image, sends it to Codex, and receives a modified bitmap. The original image and annotations are retained, while the new bitmap appears next to it without any editing controls.

The article also mentions AI HTML and AI Slides, which work the same way as AI Image: AI HTML creates a single web page element on the canvas, and AI Slides bundles multiple pages or HTML blocks into a presentation‑style sequence.

When drawing arrows, each arrow should point to a single target and each annotation should describe one action. Overlapping arrows can cause the model to mix up instructions, and text that obscures the main subject reduces recognition accuracy. The recommended workflow is to isolate one arrow at a time, keep prompts concise, and, if results are far off, zoom the target image, separate crossing arrows, shorten annotations, or provide a clean original image with explicit constraints.

Limitations include the fact that Cowart generates a new bitmap rather than preserving PSD layers or pixel‑level edit history, making it unsuitable for fine‑grained retouching. It excels at rapid versioning, side‑by‑side comparison, and conveying modification intent, but pixel‑perfect edits still require dedicated image‑editing software.

In summary, the complete workflow from installation to image modification consists of registering the marketplace, adding the plugin, opening the canvas, using AI Image (or AI HTML/Slides) to generate content, annotating for edits, and finally applying the 按标注修改 command. The key practice is to prepare clear, short prompts and annotations before invoking the model.