The History of AI

Model Context Protocol

Era 5 · Generative AI, reasoning, and agents (2022–2026) › Agents

Lineage
Acting in the world
Kind
Implementation

Who

Anthropic

What happened

Anthropic open-sources the Model Context Protocol (MCP), a standard for connecting AI assistants to the systems where data live: code repositories, business tools, databases.

Why it mattered

An open standard for connecting models to tools. OpenAI (March 2025) and Google (April 2025) adopt it, and in December 2025 it moves to a foundation under the Linux Foundation.

The full story

Every time a company wanted to connect an AI assistant to a data source, it had to write a custom integration, and every new source required another one. Anthropic proposed replacing those one-off integrations with a common protocol: whoever holds the data exposes it through an MCP server, and any AI application that speaks the protocol can connect to it as a client.

The release included the specification, software development kits, support for local servers in the Claude desktop app, and an open repository of ready-made servers for Google Drive, Slack, GitHub, Git, Postgres, and Puppeteer. Block and Apollo had integrated it, and Zed, Replit, Codeium, and Sourcegraph were working with it. The protocol borrows its ideas about message exchange from the Language Server Protocol, the standard that connects code editors with the tools for each programming language.

Adoption was rapid. In March 2025 OpenAI built it into its products, including the ChatGPT desktop app, and in April Google announced that it would adopt it. In December 2025 Anthropic donated it to the Agentic AI Foundation, a fund under the Linux Foundation that it cofounded with Block and OpenAI.

It also brought new risks: in April 2025 attacks were described that hide instructions in tool descriptions, a form of prompt injection.

On the left, three applications each joined to four data sources by twelve crisscrossing lines. On the right, the three applications and the four sources all connect to a central box, MCP, with seven lines.

Without a common protocol, three applications and four data sources need twelve integrations; with MCP, seven connections.

Under the hood

MCP defines three roles. The host application, such as Claude Desktop or a code editor, contains one client for each server it connects to, and each server exposes a particular system.

Messages follow JSON-RPC 2.0, the same format as the Language Server Protocol. A server offers three kinds of things: tools, which the model can call, such as searching a repository; resources, data the application can read, such as a file; and prompts, ready-made request templates. When they connect, client and server negotiate what each one supports.

The server can run on the same computer, communicating through standard input and output, or on another one, over HTTP.

This turns a problem of M applications times N sources, each with its own integration, into one of M plus N: each application implements the client once and each source implements the server once.

Leads to

  • Qwen3 (April 29, 2025), ChatGPT agent (July 17, 2025)

    Documented direct influence

    Qwen3 supports the Model Context Protocol natively, and ChatGPT agent connects to outside apps through connectors

Threads that pass through here

Sources

How to cite

“Model Context Protocol.” The History of AI, revised November 13, 2026. https://thehistoryofai.com/en/milestone/model-context-protocol

This entry was last reviewed on November 13, 2026.

Open in the interactive timeline