Model Context Protocol (MCP)

advancedwidespread

An open standard that provides a unified interface for AI assistants and agents to access external programs, data, and services.

In plain terms

MCP is most often compared to the USB-C connector. In the past, every device needed a separate charger and cable; today, a single connector serves almost everything. MCP gives the AI the same: a single common "socket" through which it can connect to your calendar, email, a database, or any other program.

A language model alone can only write text: it cannot see your files or book appointments for you. For an agent to actually act, it needs tools. Before MCP, a separate, custom connection had to be written for every AI application and every service.

With MCP, you only need to write the connector, the so-called MCP server, once, and any AI application that knows the standard can use it. This is why thousands of ready-made connectors for the most diverse services have appeared in a short time.

How it works

There are three players: the AI application you use (for example, a chat program or coding agent), the MCP client running inside it, and the MCP server, which represents an external service or data source. An application can connect to multiple servers at once.
Upon connection, the client queries what the server offers. The server lists its tools (for example, 'search email', 'create event') and describes what each is for and what data it expects. Alongside tools, it can also offer readable data (for example, files) and ready-made question templates.
When you make a request, the model decides based on these descriptions whether a tool is needed, and if so, which one.
The application sends a standard-format request to the server, which performs the operation and returns the result.
The model incorporates the result into its response or calls another tool. Before more sensitive operations, the application typically requests your permission.

An everyday example

In your chatbot settings (for example, in the 'Connectors' menu), you link the AI to your calendar and email. Then it is enough to write: 'check when I am free next week and write a reply to Kata' – the AI accesses both services in the background, typically via MCP.

Who is affected

developerWrite an MCP server once, and your service becomes accessible to every standards-compliant AI application. → Code

entrepreneur, executiveMCP greatly influences how easily and under what security conditions AI agents can be connected to corporate systems.

employeeThis is what makes your AI assistant a helper that can perform tasks within your workplace tools. → Work & HR

everyoneIf you connect AI to your accounts, you decide what access you grant it, so you are also responsible for part of the risks.

A common misconception

MythMCP is either a new type of artificial intelligence or a specific program that must be downloaded.

In realityMCP is only a common rule set for how AI and external programs talk to each other. By itself, it makes nothing smarter; it only enables connection.

Limits and risks

An MCP server can access your accounts, and a server running on your own machine can even access your files, so do not install connectors from unknown sources.
Hidden instructions can be inserted into text from external sources (website, email, document) that try to make the agent do something you did not request.
Too many connected tools can degrade performance: the model finds it harder to choose among them, and their descriptions take up space in the context window, that is, the amount of text the model can consider at once.
The standard only unifies the connection; it does not guarantee that the AI uses the tool well or that the connector is reliable.

Similar, but not the same

API – An API is a specific, unique programming interface for a given service, while the MCP is a unified layer designed for AI, which often builds on such APIs.
Tool use (function calling) – Tool use is the model's ability to call external functions, while MCP is the common standard that describes how to offer these tools to it.
Agent2Agent protokoll (A2A) – A2A regulates how two agents communicate with each other, while MCP regulates how an agent accesses tools and data.

Related concepts

Broader

AI-agent, Open standard, Interoperability

Narrower

MCP-szerver, MCP-kliens, Connector

On the edge

Tool use, API, Agent2Agent protokoll (A2A), Prompt injection, Kontextusablak

A brief history

Anthropic released the MCP as an open standard in November 2024. In 2025, other major AI developers, including OpenAI, Google, and Microsoft, adopted it. In December 2025, Anthropic transferred the standard to the Agentic AI Foundation, established under the Linux Foundation, so it is no longer maintained by a single company.

Where things stand

By the end of 2026, MCP has become the established industry standard for connecting agents with external tools, supported by nearly all major AI applications and developer tools. The focus has shifted to security, permission management, and corporate oversight.

Status: 2026 November

In the news

Appears in 36 Agent Age stories; the curve shows weekly occurrence in the news magazine.

AI coding agents are installing MCP servers that will never be patchedNov 3, 2026 · Code
New GEO and AI visibility tools launch: OptimizeGEO, Keytomic, Heading, Yext–Flamel.aiNov 2, 2026 · Marketing
AWS Transform can be used in GitHub Copilot via MCP for .NET projectsOct 27, 2026 · Code
Shopify, Meta–Sierra, Cloudflare and UCP: online stores open up to AI agentsOct 19, 2026 · Marketing
TikTok builds an agentic shopping assistant, direct checkout and an AI ad networkOct 19, 2026 · Marketing
Claude Code and Codex CLI updates: compression, MCP tasks, mid-reply steeringOct 19, 2026 · Code
Gemini Enterprise renders MCP Apps in chat as Google publishes developer guideOct 19, 2026 · Code
Copilot administration: model and MCP controls, server-side settings, organization billingOct 12, 2026 · Code

Also known as

MCP · Model Context Protocol · modellkontextus-protokoll · MCP protokoll · MCP-szerver · MCP server · MCP connector

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.