How to Get Comments on a Cursor Canvas (2026)

By Marcin, founder of Markloop · Last verified July 2026

What shared canvases can and can't do

Credit where due. Cursor shipped shared canvases on 20 May 2026, and the sharing side is genuinely good: hit Publish, get a link, and the canvas opens in the recipient's browser. No Cursor account, no sign-in. Compare that with Claude Code artifacts, which stop dead at your claude.ai organization - Cursor's link walks straight past that wall. If all you need is for a client to see the thing, Cursor has solved it.

What it hasn't solved is the way back:

So the current state: your client opens the canvas, spots the problem in the pricing table, and tells you about it in Slack, by email, or on a call. You re-type their words into Cursor by hand. For a one-shot "take a look", fine. For sign-off on a spec, a report an expert must correct, or anything with rounds, it leaks.

The bit nobody tells you: the canvas is already a file

As of July 2026, a Cursor canvas is rendered from a standalone HTML file the agent writes into your workspace. I verified this first-hand on 5 July 2026: I had the agent build a small report canvas and checked the workspace - the source .html was sitting right there, self-contained, inline CSS, zero external references. There's no export button on the canvas toolbar, and it turns out you don't need one. The file the canvas renders from is already in your project.

That changes the question. You're not trying to get feedback on a canvas - you're trying to get feedback on an HTML file your agent wrote. And that's a solved problem. If you can't spot the file, one prompt settles it:

Save this canvas as a standalone HTML file in the project root.

How to get anchored comments on it - from inside Cursor

Markloop is a review tool for exactly this kind of document: agent-made, self-contained HTML. It connects from Cursor over MCP, so the whole loop - push, share, collect, pull - runs from your chat. Copy-pasteable version:

1 · Connect Markloop to Cursor

Add one entry to ~/.cursor/mcp.json:

{
  "mcpServers": {
    "markloop": {
      "url": "https://app.markloop.io/api/mcp"
    }
  }
}

Cursor prompts a browser sign-in, and Settings → MCP then shows Connected · 10 tools enabled. Verified 5 July 2026 - the whole thing is one config entry and one OAuth screen.

2 · Push the canvas file

In your Cursor chat:

Push this HTML to my Markloop project as a new document.

The agent uploads the file as version 1 of a new document. No manual upload, no drag-and-drop - the file never leaves your normal flow.

3 · Share the Markloop link

Links are view-only by default. Flip on "allow comments" and anyone with the link can comment - in the browser, no login, no account. Or invite reviewers by email if you want to know who said what. Either way your reviewer sees the rendered document, not your source file, and reviewers are always free.

4 · Pull the feedback back into Cursor

When comments land, back in your Cursor chat:

Pull the open comments on that document.

Each comment arrives anchored - its number, the quoted text it was left on, and the selector it points at - as structured context in your chat, not prose you re-type. The agent applies the changes locally, pushes version 2, and the comments it dealt with get tracked as addressed. Open questions carry into the next round. You can also embed questions in the document itself - "which of these two layouts?" - and reviewers answer them inline.

FAQ

Related: Claude artifact sharing options compared · How to share a Claude Code artifact outside your organization · HTML is the new Markdown - reviewing it is the missing half