Main area
Agency
ki-orchestrierung

Model Context Protocol (MCP)

Redaktion ·

The Model Context Protocol — MCP for short — just means this: a single, shared plug through which an AI model reaches external tools and data. Instead of building a separate integration for every model and every data source, both sides speak the same protocol. Anthropic introduced MCP in late 2024, and within a year it had grown into an industry standard that OpenAI, Google, Microsoft and AWS all support too.

The problem MCP solves

Picture this: you want to give an AI assistant access to your ticket system, your database and your Google Drive. Without a standard, you build three integrations — and you rebuild them for every model. Function calling, meaning a model’s ability to call functions, does look almost identical across OpenAI, Anthropic and Google at the schema level. But you still have to wire up the tools themselves per provider.

That is the classic N-times-M problem: N models times M tools gives you a flood of one-off integrations nobody wants to maintain. MCP flips it around. You build a tool server once, and any model that speaks MCP can use it. N times M turns into N plus M.

The architecture: client and server

MCP follows a client-server pattern. On one side sits the MCP client — the AI application itself, so an assistant like Claude, an IDE or an agent. On the other side sits the MCP server — a small program that exposes one concrete capability, say access to your Postgres database or your file system.

Between the two runs a connection over JSON-RPC 2.0, a simple message format where requests and responses travel back and forth as JSON. The transport can run locally over standard input (STDIO) or over HTTP when the server lives elsewhere. The blueprint is the Language Server Protocol from the developer world, which does exactly the same for code editors: wire it up once, use it everywhere.

An MCP server exposes three kinds of building block:

  • Tools — functions the model can call, such as a database query or an API call.
  • Resources — data the model can read, such as files or database records.
  • Prompts — ready-made templates the server can offer to the model.

And that is exactly where it clicks: once a tool server exists, every MCP-capable client can use it — you never write that integration a second time.

Why this matters for agent orchestration

The standard only really pays off in orchestrating multiple AI agents. An agent system often consists of several agents that split sub-tasks — one plans, others execute, one stitches it all together. Each of them needs tools: search, database access, file operations.

Without MCP you would have to hard-wire each agent’s tools individually. With MCP you attach one tool server, and every agent reaches it over the same protocol. That makes a system not just faster to build but easier to extend: add a new server, and all agents can use it — without you touching their code.

It sounds like overkill for a small script, and there it is. But the moment your system is made of many agents and many tools, that decoupling is exactly the point where the whole thing stays maintainable at all.

Who uses MCP and how mature it is

MCP caught on remarkably fast. Within a few months of launch there were over 1,000 community-built servers, and by the end of 2025 the official registry had grown to around 2,000 entries. Over the course of 2025, OpenAI, Microsoft, Google and AWS folded MCP into their own stacks, and in December 2025 Anthropic handed governance to a neutral foundation — a sign that MCP is no longer an Anthropic house standard but shared infrastructure.

The current specification is dated 25 November 2025. Among other things it brought task-based workflows (a server can track long-running work instead of replying instantly), simplified authorization over OAuth, and an extensions framework for add-ons outside the core. Worth knowing: the field moves fast, and some of these pieces are still marked experimental.

FAQ

What is the Model Context Protocol in simple terms?
MCP is an open standard through which AI models connect to external tools and data sources. Instead of rebuilding every integration for every model, the model and the tool speak the same protocol — like a universal plug.
Who created MCP?
Anthropic introduced MCP in late 2024. Over the course of 2025, OpenAI, Microsoft, Google and AWS added support, and in December 2025 governance moved to a neutral foundation.
What is the difference between an MCP client and an MCP server?
The client is the AI application that wants to use tools — an assistant, an IDE, an agent. The server is the program offering capabilities, such as database access. An agent is always the client.
Do I need MCP for a single AI tool?
For a small script with one fixed tool, MCP is overkill. Its value shows up when many models or agents need to reach many tools — then the standard spares you the whole tangle of one-off integrations.
How does MCP differ from function calling?
Function calling is a single model's ability to call functions — but it is provider-specific. MCP standardizes the tool side on top of that, so the same tool server can be used by any MCP-capable model.

Themenuebersicht

KI-OrchestrierungMCP-Workflows