An AI assistant knows what it learned in training and what you type to it. It does not know what is happening in your accounts right now. MCP is an agreed way to give it that view: the assistant connects to a service, asks it questions, and the service answers with real data.
MCP stands for Model Context Protocol. It is an open standard that Anthropic published in 2024 and that many assistants and tools now support. The service side is called an MCP server. The assistant side, the program a person uses, is the client.
What MCP changes
Before a standard existed, each connection between an assistant and a service had to be built by hand, one pair at a time. With MCP, a service builds one server, and any assistant that speaks MCP can use it. It works like a common socket: one shape, many devices.
What an MCP server offers
| Part | What it is | Example |
|---|---|---|
| Tools | Actions the assistant can call, with the inputs each one takes | "List the websites that are down" |
| Prompts | Ready-made questions or tasks that a person can pick | "Morning check" |
| Resources | Documents or data the assistant can read | A log file |
A server may offer all three or only some of them. An assistant asks the server what it has, then uses it.
How a connection works
- The person creates a key in the service and gives it to the assistant program together with the address of the server.
- The assistant asks the server which tools and prompts it has.
- When the person asks a question, the assistant picks a tool and sends the call.
- The server checks the key, runs the tool, and returns data, and the assistant turns it into an answer in plain words.
Servers on the web use HTTP, with messages in JSON-RPC, a small standard for sending a named call and receiving its result.
Safety questions to ask
- Can it change anything? A read-only server can only show data. A server that can write can also create, edit, or delete, so it needs more trust and more care.
- Who holds the key? A key opens the account. Make one key per program or person, so that a lost key can be revoked on its own.
- Where does the data go? What the server returns is read by the AI provider you chose, under that provider's terms.
- Can text inside the data give orders? A name or a note stored in your data is read by the assistant as well. A careful server marks such text as data, and a careful assistant does not obey it.
Common misunderstandings
- "MCP is an AI model." It is only the connection. The intelligence is in the assistant.
- "MCP is just an API." It is built on web technology like an API, but it describes tools that an assistant can discover and call by itself, not fixed addresses that a programmer writes code against.
- "A connected assistant acts on its own." What it can do depends on the tools the server offers and on the rights of the key. With a read-only server it can only look.
MCP and monitoring
Monitoring data suits an assistant well. It is current, it has a clear shape, and the questions are simple: what is down, which certificates expire soon, how did the month go.
UptimeWire has an MCP server on the Business and Agency plans. It is read-only: an assistant can read websites, incidents, uptime, certificates, planned work, status pages, and clients, and it cannot change anything in the account. It offers tools and ready-made questions. You create the keys yourself, one for each program or person, and revoke any of them at any time.