Anthropic publishes the Model Context Protocol

Anthropic published the Model Context Protocol, an open standard for connecting AI models to the external systems where data and tools live — content repositories, business software, development environments. It open-sourced the specification, software development kits, and a set of pre-built connectors, or “servers,” for products including Google Drive, Slack, GitHub and PostgreSQL.

The problem MCP addressed was integration sprawl: an assistant capable of using outside tools still needed a separate, custom-built connection for every data source and application it touched, which made connected systems expensive to build and hard to scale across different models and vendors. MCP proposed a single protocol that any model could speak to any compliant server, in principle replacing the fragmented, bespoke integrations that had grown up around function-calling and early agent frameworks. Anthropic said early adopters included the payments company Block and the developer-tools firms Zed, Replit, Codeium and Sourcegraph, working the protocol into their platforms alongside Claude.

The announcement arrived a month after Anthropic shipped computer use, a beta letting Claude operate a desktop directly through screenshots and cursor commands — a second, cruder route to the same goal of letting a model act on real software, and one MCP was intended to make less necessary for anything with a proper interface.

MCP’s significance rests less on the protocol’s technical content than on what followed. In March 2025, OpenAI said it would add MCP support across its products, including the Agents SDK and, in time, ChatGPT itself — a rival lab adopting a competitor’s open standard rather than building its own, an unusual outcome in a field where each frontier lab had mostly kept its tooling proprietary. Google DeepMind chief executive Demis Hassabis confirmed Gemini would support MCP the following month, calling it “rapidly becoming an open standard for the AI agentic era.” The episode became a reference point in the argument over how much of the AI stack should be open rather than closed, and MCP servers proliferated across the developer-tool ecosystem through 2025.