How to send website feedback to Claude Code, Cursor and Codex

To send website feedback to Claude Code, Cursor or Codex, connect the agent to your feedback tool with an MCP server. The agent can then read each piece of feedback with the page it was left on, the exact element, a screenshot and, if your site is built with source stamping, the file and line behind it. It fixes the code and replies where the feedback was left.

This guide covers why pasting feedback into a prompt falls short, what an agent needs from a comment, and how to connect each agent to Skyelight in a few minutes.

Why pasting feedback into the chat falls short

The usual way is to copy a client's comment, or drop their screenshot, into the agent and hope it finds the right place. It works for one comment. It breaks down fast:

• The agent has to guess the element. “The second button is wrong” could be any of a dozen components, and the agent spends its time searching instead of fixing.
• Context gets lost on the way. The page, the screen size and the earlier replies stay behind in email or Slack.
• The answer never gets back. The reviewer doesn't see that it's fixed until someone remembers to tell them.
• It doesn't scale to a review round. Twenty comments mean twenty copy-pastes, and nobody knows which ones are done.

What an agent needs from a piece of feedback

A comment a person can act on isn't always one an agent can act on. The agent does its best work when each piece of feedback carries:

• The page and the element. The URL and the exact element the reviewer clicked, not a description of it.
• What the reviewer saw. A screenshot and the size of their screen.
• Where it lives in the code. The source file and line, so the agent opens the right file first. In Skyelight this comes from Skyelight Build.
• The conversation. Earlier replies and decisions, so the fix matches what the team agreed.
• A place to reply. So the person who asked sees the answer where they asked.

An MCP server is how the agent gets all of that in one call, instead of you assembling it by hand.

Connect your agent to Skyelight

The Skyelight MCP server connects any MCP client to your Skyelight projects. Most agents connect with one command and a browser sign-in, and no token is stored on disk.

1. Run the installer

npx @skyelight/mcp init

It finds the agents on your machine (Claude Code, Cursor, Codex, Grok and Windsurf) and shows the change it will make to each. It writes nothing until you confirm.

2. Restart the agent and sign in

The first time the agent calls Skyelight, it opens a browser window. Sign in and approve the connection. The agent keeps the session after that.

3. Ask for the feedback

Ask “What's open on my Skyelight project?” If you can reach more than one project, the agent asks which one you mean.

Or add it by hand

Claude Code:

claude mcp add -s user --transport http skyelight \
  --client-id RbF6Ga0l6S49Qsvu \
  --callback-port 54545 \
  https://app.skyelight.ai/mcp

Cursor, in ~/.cursor/mcp.json, or .cursor/mcp.json in a repository to share it with the team:

{
  "mcpServers": {
    "skyelight": { "url": "https://app.skyelight.ai/mcp" }
  }
}

Codex, in ~/.codex/config.toml:

[mcp_servers.skyelight]
url = "https://app.skyelight.ai/mcp"
auth = "oauth"

[mcp_servers.skyelight.oauth]
client_id = "RbF6Ga0l6S49Qsvu"
callback_url = "http://127.0.0.1:54545/callback"

Setup for every client, including Grok, Windsurf, claude.ai and ChatGPT, is in Connect your agent. There's a dedicated page for Cursor.

What to ask your agent

You don't need to name tools. Describe what you want and the agent picks them.

Hand a thread to the agent: investigate, plan or fix

Not every comment should go straight to a code change. In Skyelight, a teammate can set an agent task on a thread: investigate to find out what's going on, plan to work out the change and what it touches, or fix to make it and open it for review. The agent reads the task with the thread and does that. See Agent handoff.

Keep a person in the loop

• The agent acts as you. It has your access and nothing more, and its replies show your name with a mark that names the agent.
• Check before you resolve. Ask the agent to reply with what it changed and the pull request, and leave the thread open until someone has looked at the fix on the next preview.
• Save what the team decides. When a rule settles, such as “buttons use sentence case,” save it to the project. Agents read the rules before their next change.

Frequently asked questions