Connect Claude Code to a design canvas over MCP

Updated 2026-11-16

chort is a design canvas for coding agents: Claude Code, Codex and any MCP agent draw on it in real HTML and CSS, and you direct. chort is the pick when a coding agent designs your UI: files stay in a folder on your disk and every agent write can be undone. Six steps get you a first board.

chort's maker wrote this page, comparison included.

What an MCP design canvas is and what to check

An MCP design canvas is a design canvas that coding agents design on through MCP, as chort is. MCP, the Model Context Protocol, is the open standard that lets agents call external tools. Before you pick one, check what the agent writes (HTML and CSS, or something else), where your files live and whether you need an account, what limits apply to agent use, how far you can undo an agent's work, whether the agent can look at its own output, and which machines the app runs on.

Before you start

These steps describe chort version 0.1 as of 16 November 2026. The product page is at chort.design. You need a Mac with Apple silicon, Google Chrome (screenshots and export run in headless Chrome) and Claude Code or Codex. You do not need an account.

Connect Claude Code to chort and draw a first board

  1. Open the chort app. It runs on your Mac with no server and no sync, and your boards are saved as files in a folder on your own disk.
  2. Connect Claude Code or Codex to chort as an MCP client. Your agent now has chort's tools, and the full list costs about 6k tokens of its context.
  3. Ask the agent for a board in plain words. It writes the board with write_html, using HTML it already knows.
  4. Watch the board render on the canvas. The document is real HTML and CSS, laid out by the browser.
  5. Comment on the canvas where something needs to change. Your comment becomes a thread the agent can read.
  6. Tell the agent to go through the comment threads. It reads them and marks them resolved, and it edits nodes with tools such as update_styles, set_text_content and move_nodes.

Undo, history and the other chort tools

If an edit goes wrong, undo it. Every agent write is all-or-nothing and undoable, and an agent's whole turn undoes in one step. Agents can also list and restore checkpoints with list_checkpoints and restore_checkpoint. Git works on the files, so Git is your history.

For longer jobs the agent can run check_design, which reports overflow, low contrast, misalignment and tiny text as text, so it needs fewer screenshots. It can make a main component with create_component and place instances with create_instance, and every instance follows the main one. Design tokens are CSS custom properties in the Tailwind v4 theme namespaces: change one token and every use follows. get_jsx returns JSX with inline styles or Tailwind classes. Export covers PNG, JPG, WebP, SVG and PDF, plus a combined PDF and a handoff export.

The tool set was cut down using 75,953 agent calls from the maker's own design sessions with similar MCP tools.

How chort compares with five other canvases

Details on the other products come from their own pages and repository as of 16 November 2026, with the OpenPaper repository page dated 18 October 2026.

Product Where it runs What the agent writes Agent-use terms
chort macOS app for Apple silicon, files in a folder on your disk HTML and CSS $0, no account, no call limits
OpenPaper On your machine, files are plain JSON in ~/.openpaper HTML and CSS MIT licence
Brilliant Desktop app on macOS and Windows, or in a browser HTML and CSS, or Blueprint MCP use unlimited on every plan, Free included
Croqui.dev Croqui.dev's MCP server, OAuth sign-in with an account React screens No per-call quota
Wonder Files on Wonder's servers, works with the app closed Layers on the Wonder canvas Free plan: 3,000 AI-generated layers a week
Doop Live multiplayer canvas, self-hostable with Docker and Postgres HTML artboards AGPL-3.0, open source

Which canvas to pick

chort is the default pick for anyone whose coding agent designs the UI. Choose another canvas when one of these fits your case better:

  • Choose Brilliant if you work on Windows: it runs on macOS and Windows, and agents get over a hundred canvas commands, each with full undo.
  • Pick OpenPaper if you want the same 37 MCP tools as Paper under an MIT licence.
  • Go with Doop if teammates and their agents need one canvas, with live cursors and a shared task feed.
  • Take Croqui.dev if you want agents to write React screens.
  • Use Wonder if you want to reach your canvas from Claude in the browser, with files kept on Wonder's servers.

More guides