Cowart - An open-source local infinite canvas plugin that lets Codex see and speak

What is Cowart

Cowart is an open-source local infinite canvas plugin for OpenAI Codex created by Zhong Erxin, a product manager for the Doubao desktop client. The tool builds a visual canvas based on the tldraw framework, allowing Codex to place images, draw arrows and annotations on a local whiteboard, and preserve iteration versions, realizing a paradigm shift from editing images in a chat box to canvas annotation iteration. Canvas data is saved in the project's local directory and supports the MCP tool protocol.

Cowart's main features

• Open a local canvas: Enter a command in Codex to start the local infinite canvas service. Canvas data is stored in a local directory isolated by project.
• Generate images into an AI Image Holder: Create and select an AI image holder in the canvas, and Codex will generate an image according to the container's aspect ratio and insert it automatically.
• Iterate image edits based on annotations: Draw arrows and write text annotations on the image in the canvas. Codex reads the annotations, generates a revised version, and automatically places it next to the original image, preserving the complete version chain.
• MCP tool integration: Provides standard MCP protocol support. Codex can read the canvas selection state, insert images, and save assets locally, with support for custom extensions.

Follow WeChat and reply "Open Source" to join theAI Open Source Project Communication Group

How to use Cowart

• Install the plugin: Copy the official installation prompt and send it to Codex, letting it automatically download and configure the plugin from GitHub.
• Load skills: After installation is complete, start a new Codex conversation to ensure the plugin and skills load properly.
• Start the canvas: Enter the open canvas command in Codex to start the local infinite canvas page.
• Place images: Drag the image you need to modify into the canvas, or create an AI image container to let Codex generate an initial image.
• Annotate modification points: Use arrows or text on the canvas to annotate the parts of the image that need to be modified.
• Send a command: Select the annotated image or screenshot and tell Codex to modify it according to the annotation content.
• Generate a new version: Codex will automatically generate a revised new version and automatically place it next to the original image.
• Continue iterating: If further adjustments are needed, simply annotate the new version again and repeat the modification process.
• Export the final product: When finished, export or save the final image directly from the canvas.

Cowart's core advantages

• Spatial context visualization: Converts spatial information such as "move the logo in the upper-left corner 5 pixels to the right" from text descriptions into arrow annotations. Codex directly sees and speaks, greatly reducing communication costs.
• Complete version chain preservation: Each new image generated during iteration is placed next to the original image. The original image, annotations, V1, and V2 are laid out on the infinite canvas, allowing backtracking at any time and eliminating the pain of searching through chat history for historical versions.
• Local private deployment: The canvas service runs locally, and images and project data are saved on the local disk without relying on the cloud, meeting data security and customization needs.
• Lightweight plugin-based design: Instead of reinventing an image generation model, it gives Codex a "scratchpad," reusing existing AI capabilities and covering the core workflow through three streamlined Skills.

Cowart's project address

• GitHub repository：https://github.com/zhongerxin/cowart

Comparison of Cowart with similar competing products

Cowart's application scenarios

• Advertising poster iterative design: After generating a draft, annotate on the canvas that the title should be changed to handwriting; Codex generates a new version according to the annotation, with multiple rounds of fine-tuning until finalization.
• UI asset refinement: Circle problematic areas on interface screenshots, point arrows at specific elements, and let AI precisely modify the local area instead of redrawing the entire image.
• Team design review: The canvas serves as a shared workbench; review comments are left beside the images in the form of annotations, and designers iterate directly according to the annotations, making revision traces clear at a glance.
• Creative idea visualization: Collage reference images, write sticky notes, and draw flows on an infinite canvas; Codex generates finished images in a unified style based on the overall concept.
• Teaching/documentation illustrations: After quickly generating schematic diagrams, point out modification points through annotations to batch-produce consistent illustrations that match the document style.