Cursor MCP: Put Your Customers in the Editor

Cursor MCP is one url entry in mcp.json, and nearly every server you can add brings your editor one more developer tool. This one brings your customers. Point Cursor at your website AI agent and every question people asked sits next to the code that should have answered it, with no second tab and no pasted transcript. And when a bug report arrives while you sleep, the agent on your site can file it for you.

What can Cursor MCP do for your business?

Connected to your AI agent, Cursor MCP puts what customers ask next to the code that should answer them. A support tool knows what people asked, and your editor knows why they had to ask. Neither can fix it alone. In one agent chat, Cursor can read the question, find the doc or code behind it and propose the change.

Take Trailhead Maps, a fictional hiking app run by two people, with an Ultimo Bots agent answering on the website from the docs. On Tuesday Alex, one of the founders, is working on the export feature and asks in the agent chat: "Read this week's Ultimo Bots conversations. What do people ask about export, and what do our docs get wrong?"

Cursor reads the chats with list_conversations and get_conversation: nine visitors asked whether GPX export is available, and the docs never mention that it needs the Pro plan. It opens docs/export.md, proposes the line, and Alex accepts it. After the deploy, Alex asks Cursor to refresh that page, and refresh_knowledge_sources brings the new text into the agent. The next visitor gets the right answer that same day.

• Before a release: "Which questions came up about this feature since the last release?"
• After a change: "I changed the pricing page. Refresh that source and tell me when it is done."
• On a new site: "Fetch the embed code with get_embed_instructions and add it to the site-wide layout."
• Weekly: "Show me the new leads and the figures from get_insights for the last seven days."

Conversations hold what strangers typed, so read a proposed change before you accept it, as with any text the agent found. For a first run, click Clear all on the consent page and tick only the read permissions. When you want Cursor out, Revoke access on the portal's MCP connections page cuts it off at once.

What is Cursor mcp.json?

mcp.json is the file where Cursor lists its MCP servers: ~/.cursor/mcp.json for all your projects and .cursor/mcp.json for one project. MCP (Model Context Protocol) is the open standard behind it: a server lists what an AI may do, and Cursor's agent calls those tools. A remote server takes one line with its address.

{
  "mcpServers": {
    "ultimo-bots": {
      "url": "https://mcp.ultimo-bots.com/mcp"
    }
  }
}

How do I add a remote MCP server to Cursor?

To add a remote MCP server to Cursor, put its url in mcp.json, open Customize in the sidebar, and sign in. That is the whole OAuth setup: no headers, no client ID and no auth block, which Cursor only needs for static OAuth (read November 8, 2026). Cursor registers itself with the server and opens a browser tab on the Ultimo Bots consent page. Choose one workspace, untick what Cursor should not have, and click Allow connection.

The one-click link is in our Cursor guide in the help center.

Where is the Cursor MCP marketplace?

The Cursor Marketplace is at cursor.com/marketplace and in Customize in the sidebar, with official plugins and one-click install, and community servers are listed at cursor.directory. Add to Cursor on a marketplace entry installs it and signs you in with OAuth. You do not need either one: any remote server works by its URL, and Customize turns servers on and off wherever they came from.

What MCP tools work with Cursor?

Any tool a server offers over stdio, SSE or Streamable HTTP works in Cursor's agent chat. Cursor also supports MCP prompts, resources and elicitation. The server decides what is on the list, and for most servers that list is about code.

Ultimo Bots is the AI agent platform behind the chat on your site: its MCP server gives Cursor more than 100 tools for knowledge, conversations and leads, and Mo, the builder in the AI Studio, builds what the agent does, as skills on your public APIs, remote MCP servers and ready-made app actions. Cursor can only use what you approved.

No doc can say the export service is down

Docs fix the questions with a fixed answer, and Mo handles the ones that need a live system. Some of Trailhead's chats ask whether the export service is down. No doc can say that, so the agent has to check. This needs no new code in your product: skills take shape in the AI Studio, in a conversation with Mo. Alex writes one message.

Mo reads the API, asks for the key in a secure field if it needs one, builds the skill and tests it against the real endpoint, then reports what the agent can now do. Alex then adds what a two-person team really wants at the end of a failed fix: "If the service is up and the problem stays, ask which app version and device they use and what they tapped. Read it back, and after a yes open an issue in our private repo through the GitHub API. Leave out their name and email."

GitHub takes a POST to its issues endpoint with a bearer token, which is exactly the kind of public API Mo works with. Mo reads those docs, asks for the token in the same secure field, where it is stored encrypted, and has the agent confirm the details before it files anything. Mo's test chat makes real calls, so close any test issue it leaves behind.

Export goes down at 2:40 am, with both founders asleep. Twenty minutes later a hiker in another time zone asks why an export failed. The agent checks the status API, says the export service is down and gives the time of the last update, 2:40. Later that night another hiker has a failed export while everything is up. The agent walks through the fix, it does not help, and an issue lands in the repo with the version, the device and the steps. In the morning Alex opens the issue, finds the full chat in Cursor with list_conversations and get_conversation, and starts on the fix.

Your own systems work the same way when they have a public HTTPS API with an API key or bearer token, or a remote MCP server with one. Ready-made skills cover Wix, Calendly, Cal.com, Stripe, HubSpot and Mailchimp; see app actions and the AI Studio. The agent is the same on the website, a chat link and messaging channels such as WhatsApp.

What does it cost two founders?

Two founders share one inbox, and teammates are free. Working from Cursor uses no AI credits, because Cursor does the thinking on your plan with them. AI credits pay for the agent's answers and for Mo, from one balance per agent: Smart $29 a month ($19 billed yearly) with 500 AI credits and 3 skills, Boost $59 ($39) with 2,000 and 10 skills, Ultimo $249 ($199) with 10,000 and 30 skills (pricing).

Before your next commit

Ten minutes are enough to find your first wrong doc. Paste the block above into ~/.cursor/mcp.json, click Clear all on the consent page, tick only the read permissions, and ask: "What did visitors ask this week that our docs do not answer?" Fix the first gap in the commit you are already working on. Then note the first question that needs live data, because that is your message to Mo. Using other clients as well? See Claude MCP and Codex MCP.

Questions and answers