On November 11, 2026, we asked ChatGPT: “How do I give ChatGPT Codex visual feedback on a web page it built?” The answer named no tool, and its only citations were OpenAI's own help pages. A ChatGPT MCP server is the working answer to that question, and it didn't come up.
People are looking for it. In the US, Google sees about 1,300 searches a month for “codex mcp”, 880 each for “chatgpt mcp” and “chatgpt developer mode”, and 590 for “chatgpt connectors” (search data from November 2026). Those are four names for one job: getting something the model can't see into the conversation. For a web page, that something is what the reviewer saw.
What can't ChatGPT see on a page it built?
ChatGPT and Codex can write a page and change it. What they don't have is the review. The comment a client left on the pricing card on Tuesday, the element it was pinned to, the browser it was seen in, the console error that fired just before: none of that is in the repository or the chat. It sits wherever the feedback was collected.
Pasting screenshots into the chat is the manual way across. It works for one comment. MCP is the wired way, and it works for the fortieth.
What are ChatGPT connectors, developer mode and MCP?
OpenAI has renamed things since these searches became habits. This is what each term means in OpenAI's own docs, checked November 12, 2026.
MCP
OpenAI's MCP page puts it in one line: “Model Context Protocol (MCP) lets ChatGPT and Codex use tools and information from other systems.” An MCP server is the other system's side of that. It lists what the model may do, such as look up a thread or post a reply, and answers when the model asks.
Connectors
What people still search for as connectors, OpenAI's docs now call plugins. The plugin architecture page defines them: “Plugins are the packages people discover, install, share, and publish in ChatGPT and Codex.” A plugin can hold skills (written instructions), an MCP server, or both.
You can watch the rename in one address. OpenAI's admin guide at /enterprise/apps-and-connectors is now titled “Plugin controls”. It says a plugin can include apps, and that “Apps use Model Context Protocol (MCP) to connect to services.” In the API docs, the built-in connectors sit under the heading Legacy connectors, with connector_id deprecated for models released after September 1, 2026. So the thing you install is a plugin, and the part that talks to another service is an MCP server.
Developer mode
Developer mode is about adding a server of your own, one that isn't in the directory. OpenAI's current steps for that don't use the words at all: open ChatGPT Plugins, select the plus button, choose “Add custom MCP server”, enter the server's address, and finish with “Create as a plugin”.
The term survives in workspace settings. The Plugin controls guide lists a member permission, “Create plugins with MCPs”, described as “Use developer mode to create custom app connections.” The same steps page says “Account and workspace policies apply to adding and using custom MCP servers,” and OpenAI's developer pages don't list which plans that covers. In a company workspace, the person to ask is the admin.
If the tool you want is already in the plugin directory, you need none of this.
ChatGPT MCP setup: how do I add Skyelight?
Skyelight is a website feedback tool: reviewers click an element on a live page and leave a comment on it. It's in ChatGPT's plugin directory, so there is no server address to paste and nothing to switch on first.
- Open Skyelight in ChatGPT. Use the directory listing, or search for Skyelight in ChatGPT's plugins.
- Add it and sign in. ChatGPT opens Skyelight so you can sign in and approve the connection.
- Ask about your feedback. For example, “What feedback is still open on my Skyelight projects?”
ChatGPT asks you to confirm before anything that changes a thread, such as a reply or a status change. If you added Skyelight earlier as a custom connector, remove that copy so ChatGPT doesn't see two. You need an owner, admin or collaborator role in the workspace; a reviewer can't use agent tools. The rest is in Connect your agent.
How do I add a Codex MCP server for the same feedback?
ChatGPT on the web and Codex on your machine are set up separately. OpenAI's MCP page says so directly: “ChatGPT web doesn't read local Codex configuration files.” For Codex, Skyelight's installer does it in one command:
npx @skyelight/mcp init --client codexIt shows the change it will make and writes nothing until you confirm. Restart Codex, and the first time it calls Skyelight a browser window opens for you to sign in and approve. No token is written to disk.
To edit the file yourself, add this to ~/.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"That is the format OpenAI documents: a [mcp_servers.<name>] table with a url, and an oauth table under it. Two details from the same page are worth knowing. “The ChatGPT desktop app, Codex CLI, and IDE extension share this configuration,” so you set it up once. And a .codex/config.toml inside a repository scopes a server to that project, for trusted projects only. In the Codex terminal, /mcp lists the servers that are active.
Working in Claude Code or Cursor instead? The setup for those is in how to send website feedback to a coding agent.