How it works
Anthropic announced MCP as a standardized way for AI models to interact with external tools. The idea is plain. An MCP server is a small program that exposes functions, and each server describes what it can do with a simple schema. Enable a filesystem server and the assistant gets functions such as read_file and write_file. Enable a server for a database and it can query that database.
Nothing in it is exotic. The messages are JSON-RPC, a pattern used for decades, over a local connection when the server runs on your machine. A server can also be deployed remotely over HTTP, and then it has to implement the authorization the protocol requires. Servers compose: one request can use several of them in sequence. The person or the team that operates the assistant decides which servers are enabled and what each one may reach.
Where it shows up in real work
I built Airtable MCP, an open-source server that connects Airtable bases to assistants such as Claude. It began with a small annoyance. I was restructuring database schemas and kept copying schema information between the API documentation and my project. With the server in place I asked in plain language for the feature requests with the most votes, or for the number of bug reports still open. The hardest bug was authentication. The server connected to Airtable in my tests and failed when Claude used it, because of the way the token was passed through the MCP pipeline.
What it is not
MCP is not a security layer. It provides the interface between the agent and an operation, and database permissions, application authorization and operational controls still have to enforce the boundary. It is also not a way to give a password to a model: the model receives a tool definition and the minimum result it needs to answer, and the credentials stay in the secret facility of the runtime.
When you need it
Use MCP when an assistant or an agent should work with a system directly and people are still copying data in and out of a chat window. For company data, begin with specific read tools, such as one that returns the status of an order, and define their inputs, returned fields and failure behavior. A general SQL tool needs a much stronger design for isolation and query control. An instruction that asks the model to behave is not enough.