Your AI coding agent can write the code — but it can't see your
design.So we keep describing visual things in long paragraphs and hoping it understands. That's
backwards.Here's a tool that flips it 🧵
Cowart is a local, infinite canvas for Codex — built on
@tldraw.You open a real visual canvas next to your project, and the AI agent works inside it with you.
The killer feature: annotation-based editing.Screenshot your UI → scribble what's wrong right on top → Codex generates a clean, fixed version beside the
original.You point. It draws. 👇
Need an asset? Drop an "AI image holder" on the canvas, select it, ask Codex to fill it.Generated images save straight into your project's /canvas folder. Local-first, no lock-in.
Under the hood it's MCP: Codex can read what you've selected, insert images, and manage local files — so the canvas and the agent share one context.
Open source, runs locally. If you live in Codex, this turns it from a text box into a visual workspace.⭐ Try it / star it:
github.com/zhongerxin/cowart