MCP
(standardising the hands)
MCP stands for Model Context Protocol. It is a shared standard for connecting tools to LLMs. Before MCP, every team wrote custom code to plug Gmail into Claude, Slack into ChatGPT, Notion into Gemini. Every connection was bespoke. MCP is the USB-C of AI tools: one plug, works with any compatible model, built once by whoever owns the tool.
MCP turned on. The tool shelf now accepts any standard-compliant connector.
The process
An MCP server is a small program that exposes tools using a standard format. The program might sit on your laptop, on a company server, or in the cloud. It declares which tools it offers: “I provide search_email and send_email and mark_read.” It speaks the standard protocol.
An MCP client lives inside an AI app like Claude Desktop, Cursor, or Claude Code. The client discovers which MCP servers the user has connected. The client asks each server for its tool list. All those tools get added to the LLM’s available tool set.
A user can add or remove MCP servers without any custom code. The AI app supports the protocol. The servers speak the protocol. They interoperate.
You’ve encountered this when…
You connected Gmail, Google Calendar, or a third-party service to Claude Desktop or Cursor, and it worked without you writing any code. That is MCP under the hood. The app you were using speaks MCP. The service speaks MCP. They plugged into each other on their own.
A familiar example
Think about plugging a USB-C cable into any modern laptop, phone, or display. Ten years ago, every device had its own cable shape. Now one shape fits them all. MCP is the same move for AI tools. A team builds one MCP server for their product, and every MCP-compatible AI app can use it on day one.
Variants include
MCP servers you run locally
A server that runs on your own machine, exposing tools that touch your filesystem, your local database, or your local apps. Private by default. Nothing leaves your laptop unless you want it to.
MCP servers run by a vendor
A server run by a company, exposing their product’s tools over the internet. Asana, Linear, GitHub, and many others offer hosted MCP servers. You connect once with a login.
The open directory
A public catalog of available MCP servers, from official vendor servers to community projects. The ecosystem is young, so quality varies. Some servers are polished. Others are half-built hobby projects.
The breaking point
MCP is less than a year old. The ecosystem is half-built. Some connectors are excellent. Others are half-maintained hobby projects. Authentication, permissions, and trust are still being worked out across the industry. This situation will change fast over the next 18 months.
Your takeaway
The next time you connect an AI app to an external service and it works, MCP is why. The standard is the reason the ecosystem started snapping together instead of fragmenting further.