How do I give an AI coding agent feedback on my website?

Pick the method by the feedback. Use a screenshot for one visual fix. Use a bug-report tool with an MCP server for defects. Use Capture by Display to comment on the text of a page and give the agent one link with all comments.

Can Cursor or Claude Code read Marker.io or BugHerd reports?

Yes. Both vendors have MCP servers that list Claude Code and Cursor as clients. The agent reads the report, the screenshot, and the page details.

What is Capture by Display?

A Chrome extension. It copies the page you see, lets you comment on its text, and publishes the copy with your comments to a Display link.

Does Capture by Display need an account?

Yes. You sign in to Display to publish. Comments are optional.

Does the agent change my website directly from Capture?

No. Capture publishes a copy of the page. Your agent reads the copy and the comments, then makes the change in your repository when you ask it to.

Can a client comment on a captured page?

Not yet. Invited people outside your organization can view the copy. Members of your organization can comment. For client feedback on a staging site, a widget tool fits better today.

Can I comment on an image or a button?

Not in Capture yet. Comments attach to selected text. Select the nearest label or heading and describe the element.

How to Give an AI Coding Agent Feedback on a Web Page

Display is our product. We checked the other vendors' pages on 16 October 2026.

The short answer

Pick the method by the kind of feedback.

Your feedback	Best method	What the agent gets
One small visual fix	A screenshot in the agent's chat	A picture, with no text or structure
"Make it look like this other page"	The URL	The page's HTML, but not your intent
A bug with errors or broken requests	A bug-report tool with an MCP server	The report, a screenshot, console logs, and network requests
Copy and design changes across one page	Capture by Display	A copy of the page with each comment attached to its text

The methods do not compete. A team often uses a bug tool for defects and comments on a page copy for copy and design changes.

Method 1: a screenshot in the chat

Cursor, Codex, and Claude Code accept images. You take a screenshot, draw an arrow, and type the change.

Good for: one quick fix that you can see.

Problems:

The agent reads pixels. It must guess which component and which line of code made them.
Several screenshots in one message lose their order and their context.
Your teammates do not see the feedback. It stays in your chat.
Long pages need several screenshots, and the agent cannot scroll them.

Method 2: give the agent the URL

Many agents can read a page or open it in a browser. You paste the URL and describe the change in words.

Good for: "copy the layout of this page" and questions about public pages.

Problems:

The agent sees the page, but it does not see which sentence you mean. You describe the location in prose, and the agent may pick the wrong element.
Pages behind a login, or pages that load content as you scroll, can look different to the agent.
A list of ten changes in one message is hard to check off one by one.

Method 3: bug-report tools with MCP servers

Website feedback tools collect reports from a page, often from clients or testers. Three of them now give agents direct access through MCP servers.

Marker.io. A widget or browser extension creates reports with screenshots and annotations. Marker.io says its MCP server gives the agent the screenshots, URL, console logs, network requests, browser and OS details, and the reporter's description. It lists Claude, Cursor, Windsurf, and other MCP clients. Marker.io includes the MCP server on every plan. Starter costs $39 a month with annual billing, for 3 users. Reporters are free and unlimited.

BugHerd. Reviewers pin feedback to elements of a live site. BugHerd's MCP server uses OAuth and works with Claude Code, Cursor, ChatGPT, Codex CLI, and other clients. Per task, it shares the description, screenshot, page URL, and the CSS selector of the element. The agent can also update tasks and add comments. Standard costs $42 a month with annual billing, for 5 members, and lists MCP among its features.

Jam. A Chrome extension records a bug with video, console logs, network requests, and user events. Jam's MCP server lets an agent open a Jam link and load that context. Jam lists ChatGPT, Claude, Cursor, VS Code, Windsurf, and OpenCode. Jam's MCP page does not mention a plan requirement.

Prices and features checked on marker.io/pricing, bugherd.com/pricing, and the vendors' MCP pages on 16 October 2026.

Good for: bugs. The technical context, such as console errors and failed requests, is exactly what a coding agent needs to find a defect.

Less good for: copy and design reviews. A design review has many small comments on text and layout. Each one becomes a ticket.

Method 4: comment on the page with Capture by Display

Capture by Display is a Chrome extension for desktop Chrome. It makes an HTML copy of the page you see. You can comment on its text first. Then it publishes the copy, with your comments, to a Display link. You need a Display account.

How it works:

Open the page in desktop Chrome and select the Capture icon.
Sign in to your Display account.
Select text on the page and select Comment. Write what should change. Repeat for each change.
Open the Publish panel. Capture prepares the page and shows a preview with any missing parts.
Choose who can open the copy, then publish.

The result is an ordinary Display artifact. Your teammates open the link and read each comment next to its text. The agent reads the same link.

Hand it to the agent. Connect your agent to Display through MCP once. Then open the artifact, open More, and select Copy edit prompt. Paste the prompt into Cursor, Codex, or Claude Code. Or write your own:

Read the Display artifact at <link> and its open comments.
Each comment marks a change to our website.
Make the changes in the HTML copy first and publish a new version at the same link.
After I approve that version, apply the same changes in our repository.
List any comment you did not address.

This order lets the team review the change on a page before anybody edits the repository. The agent works on the copy first, and people check the result at the same link.

What Capture keeps: the rendered text, layout, inline SVG, and styles. It embeds public images and stylesheets when you allow access to their sites.

What Capture does not do:

It is not the site's source code. The copy shows what the page looked like. The agent still needs your repository to make the real change.
It comments on text only. You cannot pin a comment to an image or a button yet. Select the nearest heading or label instead.
It skips some content. Cross-origin frames, canvas and WebGL pictures, protected video, and private images may be missing. Capture lists what it left out.
It runs in desktop Chrome only.
It has size limits. A captured page can be up to 10 MB on the Free plan and 50 MB on paid plans.

Who can comment. Members of your organization can comment on the published copy. Invited people outside your organization can view it but cannot comment yet. Capture publishes as Private by default. You can choose Company, Public, or specific email addresses.

Display is the docs workspace for agents and people. Any agent publishes docs and files to a link. Your team opens the link with the account they already use. They comment and edit in-line. Any agent or teammate can publish the next version to the same link.

Which method to use

One fix, right now: a screenshot or a sentence in the chat.
A defect with errors: a bug-report tool with an MCP server, such as Marker.io, BugHerd, or Jam.
Client feedback on a staging site: a widget tool such as Marker.io or BugHerd, because reporters need no account there.
A copy or design review by your team: Capture by Display, then let the agent read the comments.
A review of a competitor's page: Capture by Display. Read how to review a competitor's website with your team.

For the steps of a review loop with an agent, read how to review AI agent output as a team. For other ways to save a page, read how to save a web page as HTML and share it.

Sources