Connect Cursor, Claude, and ChatGPT

Paste one address, sign in on chat.sh, and choose what Cursor, Claude, or ChatGPT may do in your help center.

Cursor, Claude, and ChatGPT can read your help center and write in it. You paste one address, sign in, and choose what the app may do. There is no API key.

The same steps work for every workspace you allow. A connection belongs to you, so one app can cover several workspaces.

Before you start

  1. Open a workspace that has a lifetime plan. An app cannot connect to a workspace that does not.
  2. In the sidebar, open AI agents.
  3. Copy the server address. It is https://chat.sh/mcp.

Paste that address only. Do not paste an access token, a client id, or a client secret. chat.sh asks you to sign in, then shows you the workspaces and the permissions.

If chat.sh says the app registered itself and has not been verified, continue only when you started this from an app you trust.

What you can allow

You choose this on the sign-in page the app opens. Reading stays on. The rest you can leave off.

  • Read articles and collections. The app can list your collections and read article bodies, including drafts.
  • Create drafts and edit articles. A new article starts as a draft and uses none of the workspace's pages until it is published. The app can also add a collection. Editing an article that is already published changes the live page at once. A longer body is refused if it would go past the workspace's pages. Images the app adds do not use pages.
  • Publish and unpublish. This puts an article on the public help center, or takes it back to a draft. Its pages count against the workspace from the moment it is published. See AI credits and pages.
  • Delete articles. This is permanent. The app has to repeat the article's title exactly, or nothing is deleted. Leave this off unless you want the app to be able to delete.

You can revoke an app later. That signs it out of every workspace at once.

Cursor

Cursor signs in with the address above. Do not add an Authorization header.

  1. Open Cursor Settings, then Tools & MCP.
  2. Choose New MCP server, or open the chat.sh entry and select Connect.
  3. You can also add this to ~/.cursor/mcp.json, or to .cursor/mcp.json in one project:
{
  "mcpServers": {
    "chat.sh": {
      "url": "https://chat.sh/mcp"
    }
  }
}
  1. Sign in on chat.sh when the browser opens. Choose the workspaces and what Cursor may do, then select Allow.

Ask it to list your articles. Cursor asks before it runs a tool. Connect it again if you revoke it.

Claude

Claude reaches chat.sh from Anthropic's side, so the address has to be the public one above.

On a Pro or Max plan

  1. Open Claude and go to Customize, then Connectors.
  2. Select Add custom connector.
  3. Paste https://chat.sh/mcp.
  4. Leave authentication required. Leave the client id and client secret empty, and leave Advanced settings empty.
  5. Add it, then connect. Claude sends you to chat.sh.
  6. Sign in, choose the workspaces, choose what Claude may do, and select Allow.

On a Team or Enterprise plan

An owner adds the connector before anyone else can use it.

  1. Open Organization settings, then Connectors.
  2. Select Add.
  3. Hover over Custom, then select Web.
  4. Paste https://chat.sh/mcp. Leave advanced settings empty.
  5. Select Add.

Each person then connects with their own chat.sh account, from Customize, then Connectors, and chooses their own workspaces and permissions.

Use it

Start a chat and ask Claude to work on the help center. A new article stays a draft until you say to publish it, and only if you allowed publishing.

ChatGPT

ChatGPT on the web connects to the same public address.

  1. Open ChatGPT Plugins.
  2. Select Add, then Create MCP App.
  3. Name it chat.sh. Paste https://chat.sh/mcp. Leave Connection on Server URL.
  4. Leave Authentication on OAuth. Leave Advanced OAuth settings empty.
  5. ChatGPT warns that a custom MCP server introduces risk. Tick I understand and want to continue, then select Create.
  6. Sign in on chat.sh, choose the workspaces and what it may do, and select Allow.
  7. In a chat, turn the chat.sh app on, then ask it to work on the help center. Try: "List the articles in my help center, including drafts."

Do not choose an API key or a pasted token. chat.sh signs you in.

What to ask first

The app can see the help centers you allowed. Start with one of these.

  • List the collections and articles in my help center, including drafts.
  • Write a draft explaining how customers add a custom domain. Do not publish it.
  • Find the article about inviting teammates. Show me the change before you save it.

A new article is a draft. It is not on the public site, not in search, and not used for answers until you publish it. Publishing spends pages from the workspace's allowance.

Change what an app can do

Connect it again from that app. On the page chat.sh opens, tick the permissions you want. Reading stays on. Select Allow. That replaces what the app could do before.

If connecting again does not open chat.sh, remove the app there and add https://chat.sh/mcp once more.

Stop an app

  1. Open AI agents in the sidebar.
  2. Find the app under Apps.
  3. Select Revoke.

That app has to connect again before it can do anything. Revoking signs it out of every workspace it could use, not only the one you are looking at.

If it does not connect

  • The workspace has no lifetime plan. Buy one, or pick a workspace that already has one, on the sign-in page.
  • A client id or secret was typed into advanced settings. Remove the connector and add it again with those fields empty.
  • On Claude Team or Enterprise, an admin has not added the connector for the organization yet.
  • The address is not exactly https://chat.sh/mcp.
PreviousSubmit your sitemap