What Is MCP (Model Context Protocol)? A Plain Explanation
MCP is an open standard for connecting AI applications to data, tools and workflows. What it is, how it works, who governs it and where to start.

MCP (Model Context Protocol) is an open standard for connecting AI applications to external systems: files, databases, APIs and workflows. The official docs compare it to a USB-C port: just as USB-C gives a standard way to connect devices, MCP gives a standard way to connect an AI application to data and tools (modelcontextprotocol.io).
In practice, MCP is what lets an assistant such as Claude or ChatGPT read your calendar, query a database or open a ticket, without every company writing a separate integration for every model.
Key takeaways
- MCP is an open protocol: an MCP server is written once and works with any compatible application.
- It has three parts: the host (the AI app), the client (one connection per server) and the server (which offers data and tools).
- A server exposes three things: tools, resources and prompts.
- Since December 2025 it is governed by the Agentic AI Foundation, inside the Linux Foundation.
- A local server runs with your permissions: install only what you have reviewed.
What problem does MCP solve?
MCP solves the problem of repeated integrations. Before it, every AI application needed its own connector for every data source: three assistants and ten systems meant thirty different integrations. With a common standard, each system publishes an MCP server once and any compatible application can use it.
Anthropic introduced it on 25 November 2024 as "an open standard that enables developers to build secure, two-way connections" between their data sources and AI-powered tools (Anthropic). A year later the Linux Foundation counted more than 10,000 published MCP servers and adoption by Claude, Cursor, Microsoft Copilot, Gemini, VS Code and ChatGPT (Linux Foundation, 9 Dec 2025).
How does MCP work inside?
MCP uses a client-server architecture with three participants (official architecture):
- Host: the AI application you use, for example Claude Code, Claude Desktop or Visual Studio Code.
- Client: a component inside the host. The host creates one client per server and each client keeps a dedicated connection.
- Server: the program that provides context. It can run on your computer (a local server) or on the provider's cloud (a remote server).
An example from the docs: Visual Studio Code acts as the host. When it connects to Sentry's MCP server it creates a client for that connection; when it then connects to the local filesystem server, it creates another client.
Underneath there are two layers. The data layer uses JSON-RPC 2.0 for messages: discovering capabilities, listing tools, calling them and receiving notifications. The transport layer defines how those messages travel: over standard input and output (stdio) when the server runs on the same machine, or over streamable HTTP when it is remote, with token authentication and OAuth recommended.
What does an MCP server offer?
An MCP server offers three kinds of things, which the protocol calls primitives:
| Primitive | What it is | Example |
|---|---|---|
| Tools | Functions the AI can execute | Query a database, create an issue, call an API |
| Resources | Data that gives context | A file's contents, a database schema |
| Prompts | Reusable templates for working with the model | Examples of how to use that server's tools |
The usual flow is simple: the client asks for the list of tools (tools/list), hands it to the model, and when the model decides to use one, the client executes it (tools/call) and returns the result. Clients have primitives too: with elicitation, a server can ask the user for missing information or a confirmation before it acts.
The current version of the specification is 2026-07-28. Among its changes, the protocol became stateless (every request carries its version and capabilities) and deprecated features such as sampling and roots (official architecture). If you read tutorials from 2025, check which version they were written against.
Who governs MCP?
MCP is governed today by the Agentic AI Foundation (AAIF), a directed fund inside the Linux Foundation. It was announced on 9 December 2025 with three founding projects: MCP, contributed by Anthropic; goose, Block's agent framework; and AGENTS.md, the instruction format for coding agents contributed by OpenAI. Its platinum members are Amazon Web Services, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft and OpenAI (Linux Foundation).
A neutral foundation matters for a practical reason: no single company can change the rules alone, so building on MCP carries less risk than building on one provider's closed API.
The AAIF was a partner of Hacktoberfest México 2026, the hackathon we ran in Mexico City with a focus on open AI; what the 16 teams built includes projects with agents and open models.
What should you watch for in security?
The first thing to watch is that a local MCP server runs with the same permissions as your application. According to the official security guide, a downloaded server can run commands, read files or send data out if it comes from a malicious source (security best practices).
Simple rules to start with:
- Install servers from sources you know, ideally the provider's official ones or the project's reference servers.
- Read the full command your client is about to run before you approve it.
- Grant the least access: one folder, not your whole disk; tokens with just the scope needed.
- Prefer stdio for local servers and require authentication if a local server listens over HTTP.
- If you build a server, never forward to another API a token that was not issued for your server: the specification forbids it.
How do you get started with MCP?
To start, connect an existing server to an application you already use, then write your own.
- Use one. In Claude, ChatGPT, VS Code or Cursor, add an official server (for example the filesystem one or your ticketing tool's) and ask for something that needs it.
- Inspect it. The project publishes the MCP Inspector to see which tools a server exposes and try them by hand (repository).
- Write one. Take an official SDK and expose a single useful tool from your work, for example checking an order's status.
- Build it with others. A hackathon is a good place for that: what is a hackathon explains how they work.
To hear about Mobil3's next calls and events, join our subscription list.
Sources
- Model Context Protocol, "What is the Model Context Protocol (MCP)?"
- Model Context Protocol, "Architecture overview" (version 2026-07-28)
- Model Context Protocol, "Security Best Practices"
- Anthropic, "Introducing the Model Context Protocol", 25 Nov 2024
- Linux Foundation, Agentic AI Foundation announcement, 9 Dec 2025
- MCP Inspector
Frequently asked questions
- What does MCP mean in AI?
- MCP stands for Model Context Protocol, an open standard for connecting AI applications to external data sources, tools and workflows.
- Who created MCP and who governs it today?
- Anthropic open sourced it in November 2024. Since December 2025 it is a project of the Agentic AI Foundation, a directed fund of the Linux Foundation.
- What is the difference between an MCP client and an MCP server?
- The server is the program that offers data and tools. The client lives inside the AI application (the host) and keeps a dedicated connection to one server.
- Is MCP safe?
- The protocol has security rules, but a local server runs with your permissions. Install servers only from sources you trust, read the command they run and give each one the least access it needs.
- Which applications support MCP?
- The official docs name Claude, ChatGPT, Visual Studio Code and Cursor, among others. The Linux Foundation reported more than 10,000 published MCP servers in December 2025.