You ask an AI assistant to find an invoice. It gives you instructions for searching your files. What you wanted was the invoice itself.
MCP, short for Model Context Protocol, is an open standard that lets compatible AI applications connect to external tools and data. It gives those connections a common language. Whether an assistant can read a folder or send an email still depends on the tools and permissions you provide.
This guide explains the moving parts, follows one simple example and gives you a checklist for choosing your first connection.
What MCP adds to an AI assistant
A model can work with the information placed in its context. An application can also give it tools, such as a file search or calendar lookup. MCP standardizes how an AI application discovers and communicates with capabilities supplied by another program.
Think about the difference between describing your calendar and connecting a calendar lookup. In the first case, you supply the details. In the second, a tool retrieves them from the source when asked.
The official MCP introduction describes this shared connection layer. It does not promise that every assistant supports every server or that a connection makes the answers correct.
For a concrete notes use case, our Obsidian MCP guide shows where this connection fits.
Host, client and server in plain language
The host is the AI application you use. A client inside that application handles communication with a server. The server is the program that exposes a particular set of capabilities.
A server might connect to your notes, a database or a service. Some run on your computer. Others run remotely. “Server” describes its role here, so it does not necessarily mean a separate physical machine.
The MCP architecture guide explains these roles. For a beginner, the useful question is: which program gets access to which information?
Servers can expose three main kinds of capability:
- Tools: operations such as searching invoices or creating a draft.
- Resources: information the application can read, such as a document.
- Prompts: reusable interaction templates provided by the server.
These are different parts of the protocol, described in the server concepts guide. A particular server may offer only some of them.
A small invoice example
Imagine a freelance designer who wants to find an unpaid invoice. This is an illustrative workflow, not a test of a particular product.
- The designer asks the assistant to find the latest invoice for a named client.
- The assistant uses an available search tool through an MCP connection.
- The server returns matching records from the connected folder or service.
- The assistant reports the invoice number, date and source location.
- The designer checks the result against the original invoice.
Finding an invoice and emailing a reminder are separate actions. A read-only search connection can help with the first task without providing the ability to perform the second.
If you later want draft reminders, add that capability deliberately. Decide separately whether sending requires approval. Do not assume the word “MCP” settles those choices.
Is MCP the same as an API?
An API is an interface a program can use to interact with another system. An MCP server can call a service’s API behind the scenes, then expose selected capabilities through MCP.
You can build AI tools without MCP. Its value is the shared interface, which can reduce custom connection work across compatible applications. It does not replace every service’s API or remove its authentication rules.
Check these five things before connecting
- Purpose: What exact task should this connection make easier?
- Source: Who maintains the server, and where is its official documentation?
- Access: Can it read, write, delete or send? Which folders or accounts are included?
- Data flow: Does it run locally or remotely, and what leaves your device?
- Control: How do you review actions and revoke the connection?
Retrieved content can contain misleading instructions. A document’s contents should be treated as information to inspect, not permission to change accounts or send files. Application safeguards and your access settings still matter.
Our guide to AI agent boundaries helps separate reading, drafting and actions that need your approval.
Try one read-only task first
Use a folder of sample files that contains no private information. Connect only the access needed to search it, then try this request:
Find the document about [topic] in [approved source]. Give its title, location and the passage that supports your answer. If nothing matches, say so. Do not create, edit, delete or send anything.
Check the result yourself. Also ask for a document you know is missing. A useful connection should handle that limit honestly instead of inventing a filename.
Once that works, choose a specific next step. If your notes live in Obsidian, the Obsidian MCP guide covers that use case. Add connections when they solve a real task, one permission decision at a time.
