Cowart - an open-source local infinite canvas plugin that lets Codex interpret images. - AiBoss

What is Cowart?

Cowart is an open-source plugin for OpenAI Codex, developed by Zhong Erxin, product manager of Doubao Desktop. The tool uses the tldraw framework to build a visual canvas, allowing Codex to place images, draw arrows, and retain iterative versions on a local electronic whiteboard, achieving a paradigm shift from modifying images in chat boxes to iterative canvas annotation. Canvas data is stored in the project's local directory and supports the MCP tool protocol.

Cowart's main functions

Open the local canvasYou can start the local unlimited canvas service by entering the command in Codex. The canvas data is stored in a local directory in isolation by project.
Generate images to AI Image HolderCreate an AI image holder in the canvas and select it; Codex will generate an image proportional to the container and insert it automatically.
Iterative graph modification based on annotationDraw arrows and add text annotations to the image on 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 integrationIt provides standard MCP protocol support, and Codex can read canvas selection status, insert images, save resources locally, and supports custom extensions.

How to use Cowart

Install pluginsCopy the official installation prompts and send them to Codex so that it can automatically download and configure the plugin from GitHub.
Loading skillsAfter installation, create a new Codex dialog to ensure that plugins and skills load correctly.
Launch canvasEnter the command to open the canvas in Codex to launch the local infinite canvas page.
Add imageDrag the image you want to modify into the canvas, or create an AI image container and let Codex generate the initial image.
Mark modification pointsUse arrows or text to mark the parts of the image that need to be modified on the canvas.
Send commandSelect the labeled image or screenshot to tell Codex to modify it according to the labeled content.
Generate a new versionCodex will automatically generate a revised version and place it next to the original image.
Continuous iterationIf further adjustments are needed, simply mark the changes again in the new version and repeat the modification process.
Export finished productOnce completed, export or save the final image directly from the canvas.

Cowart's core advantages

Spatial Context VisualizationSpatial information such as shifting the logo in the top left corner 5 pixels to the right is transformed from text descriptions into arrow labels, allowing Codex to directly interpret images and significantly reduce communication costs.
Version chain fully preservedEach iteration generates a new image next to the original image. The original image, annotations, V1, and V2 are tiled on an infinite canvas, allowing for easy backtracking and eliminating the pain of searching through chat history for historical versions.
Local private deploymentThe canvas service runs locally, with images and project data stored on the local disk, without relying on the cloud, thus meeting data security and customization needs.
Plug-in lightweight designInstead of creating new graph models, Codex is given a "scratch paper" to reuse existing AI capabilities and cover the core workflow through three streamlined skills.

Cowart's project address

GitHub repositoryhttps://github.com/zhongerxin/cowart

Cowart's Competitive Product Comparison

Dimension	Cowart	Bragi Canvas
position	Codex Local Infinite Canvas Image Iteration Plugin	Obsidian Node-Based AI Workspace Generation Plugin
platform	Local web service, dependent on Codex CLI	Obsidian Desktop Plugin
Canvas type	An infinite whiteboard based on TL-DRAW, allowing you to freely place images and annotations.	A node-based canvas based on Obsidian Canvas, with connecting lines serving as prompts.
AI Integration	Call GPT Image and other models using Codex Skill + MCP	Supports 20+ providers including OpenAI, Anthropic, Gemini, and ElevenLabs.
Core Interaction	Arrow annotations + text annotations → AI iteratively modifies the image based on annotations	Select nodes + connect lines → Batch generate images/videos/text/audio
Generate type	Focus on image generation and iterative editing	Supports multimodal generation of images, videos, text, voice, music, and sound effects.
Data storage	Local project directory `canvas/` persistence	Obsidian Vault local storage, supports `.bragi` Workflow import/export
MCP support	Provides MCP tools for reading canvas state and inserting images.	Optional local MCP server exposes canvas operations to Agent
Version Management	The original and revised versions are preserved side by side, and the version chain is visualized.	The generating node automatically connects back to the source node, forming a generation link.

Cowart application scenarios

Iterative Design of Advertising PostersAfter generating the first draft, the title is marked on the canvas and changed to handwritten font. Codex generates a new version according to the annotations, and multiple rounds of fine-tuning are carried out until the final version is completed.
UI asset refinement: Select the problem area in the screenshot, point the arrow to the specific element, and let the AI accurately modify the local area instead of redrawing the entire image.
Team Design ReviewThe canvas serves as a shared workspace, where review comments are left as annotations next to the images. Designers can directly iterate based on these annotations, making the changes immediately visible.
Creative concept visualization: Collage reference images, write notes, and draw processes on an infinite canvas, and Codex generates a unified style of finished image based on the overall concept.
Teaching/Document IllustrationsAfter quickly generating schematic diagrams, you can use annotations to indicate the points to be modified and generate batches of consistent illustrations that match the document style.