AI agents, Workflow
Hand review feedback straight to your coding agent
Connect Claude Code, Cursor or Codex to Notemark over MCP, so your agent reads the comment, the element and the screenshot, then replies when the fix is in.

A comment like "this heading wraps badly on mobile" is clear to the person who wrote it. To a coding agent it is missing almost everything: which page, which element, which viewport, what it looked like. Someone usually fills that gap by hand, copying text into a prompt and attaching a screenshot.
Notemark keeps that context with every comment. Each one records the page address, the element that was clicked and the viewport, and gets a cropped screenshot of that area. The Notemark MCP server hands all of it to your agent directly.
Connect your editor
The server address is https://www.notemark.app/api/v1/mcp and it uses streamable HTTP. The quickest setup is under Settings, Connections: pick your editor and copy the configuration. Sign in with OAuth is the default, and Cursor and VS Code can install it in one click.
For Claude Code:
claude mcp add --transport http notemark https://www.notemark.app/api/v1/mcp
For Codex:
codex mcp add notemark --url https://www.notemark.app/api/v1/mcp
codex mcp login notemark
The MCP server needs the Solo plan or higher.
What the agent can read
get_canvas_contextgives an overview of a canvas: its pages, comment counts and open comments as Markdown.list_commentsfilters comments by status, label, viewport or page.get_commentreturns the full detail: the element's HTML, its selector, the screenshot and the environment it was left in.get_comment_screenshotreturns the screenshot as an image the agent can look at.
That is enough for an agent to find the component in your codebase and see the problem, instead of guessing from a sentence.
Close the loop from the editor
When the fix is in, the agent can say so where the reviewer will see it:
add_replyanswers without resolving, for questions or partial fixes.set_comment_statusmoves a comment to in progress while work is under way.resolve_commentresolves it, with an optional reply that explains the change.assign_commenthands a comment to a teammate, who gets an email.
These need the Resolve & reply scope if you connect with an API key. A key with only Read comments lets an agent look without touching anything.
Give an agent one canvas, not the workspace
For a contractor or a one-off task, open a canvas's Share panel and use Copy MCP URL. That address covers that canvas only and works without a workspace key. It stops working if the share link expires or has a password, so it ends with the review round.
A loop worth trying
- Ask your agent for the open comments on the canvas.
- Let it pick one, read the element and screenshot, and make the change.
- Have it resolve the comment with a one-line reply saying what changed.
- Reload the canvas and check the fix on the live page.
Setup details for every editor are in the MCP server guide.

