Connect Your AI Agent to Your Cloud with MCP

If you already ask an AI agent to write your code, the next question comes quickly: why can't it check the server that code runs on? Carpathian is an AI native cloud platform, which means your agent connects to your cloud directly over MCP. It can look up your servers, domains, firewall rules and deployments, propose changes, and make them once you've agreed or straight away if you've told it to, without you copying output between a terminal and a chat window.

The agent never gets a master key. Each one has its own identity and its own permissions, and those permissions can't go past what you hold yourself. Every call it makes is recorded. You choose how far each agent goes alone: it can only read, ask before every change, run changes and ask before deletes, or run everything with nobody asked, which only its owner can turn on, and confirms in a popup. New agents start by asking.

What is MCP, and why does it matter for your cloud?

MCP, the Model Context Protocol, is an open standard for connecting AI applications to the tools and data they work with. A client such as Claude Code, Claude Desktop, Cursor or VS Code speaks MCP on one side, and a service like Carpathian speaks it on the other, so any client that supports the protocol can reach any server that offers it. The specification is public at https://modelcontextprotocol.io.

For a cloud platform, that means your agent stops guessing. Ask it why a deploy failed and it reads the deployment log. Ask whether a domain points at the right server and it checks the routing rule. The answer comes from your account as it is right now, not from what the agent remembers about how clouds usually work.

How do you connect an agent?

There are two ways in, depending on where the agent runs.

• Sign in from your editor or desktop app. Add Carpathian as an MCP server in your client, and it opens a Carpathian sign-in page. You pick which agent the client acts as and confirm what it may do. The client never sees your password, and the access it gets is tied to Carpathian alone, so it can't be replayed anywhere else.
• Use an agent key for headless work. For a script, a scheduled job or an agent running on a server, create an agent key. Each key belongs to exactly one agent, and you can revoke it without touching anything else.

You manage both under Agents in the dashboard. Each agent has its own page with the keys and sign-ins connected to it, what it may do, its mode, its limits, and its latest calls.

What can an agent see and change?

An agent can do what three things allow at once: the permissions you gave that agent, the permissions of the person it's connected through, and the products your organization has turned on. All three are checked again on every call. If someone on your team loses access to a server, an agent connected through them loses it on the next request, not when somebody remembers to clean up.