The developers of the Model Context Protocol, an open-source technology that underpins many artificial intelligence applications, today released a new version of the software.

The release is described as the biggest update to the project since its launch.

The Model Context Protocol, or MCP, was open-sourced by Anthropic PBC in November 2024. The AI developer donated the project to an industry consortium called the Agentic AI Foundation about a year later. The group, which operates under the wing of the Linux Foundation, is backed by Anthropic, OpenAI Group PBC and other major tech firms.

MCP enables AI applications to interact with external systems. For example, an AI agent can use MCP to fetch a record from a database or update an online store’s product catalog. The technology also lends itself to more complex tasks.

When an MCP-powered AI application wishes to perform an action in an external system, it sends the system a request that details its requirements. It also sends metadata, or contextual information. The metadata contains details such as the version of MCP that the AI application uses.

The biggest change in today’s update is an overhaul of how MCP processes request metadata. Previously, AI applications sent metadata via a relatively complex workflow called a handshake. The new MCP release removes the workflow, which has several benefits. One is that it makes it easier to scale MCP applications in response to traffic spikes. The update also simplifies the task of recovering from server outages.

The applications that an AI model accesses via MCP often run across multiple servers. MCP’s handshake mechanism created complexity because it sent requests and their metadata to only one of those servers. If everything depends on a single server, a hardware issue can cause service disruptions. Having a single point of failure also complicates scaling.

The new version of MCP replaces the handshake mechanism with a so-called stateless protocol core. The technology packages metadata directly into MCP requests. That removes the need for an external metadata management mechanism like the handshake workflow, which avoids the associated technical challenges.

Today’s MCP update also introduces several other enhancements.

The protocol’s developers improved the authorization mechanism that it uses to regulate AI applications’ access to external systems. The mechanism, which ensures that an application can only access features it has permission to use, is now less prone to certain types of errors. It’s also less vulnerable to a type of cyberattack known as an OAuth mix-up that enables hackers to steal login credentials.

The improved authorization module is joined by a new extension framework. The latter component makes it possible to extend MCP’s core feature set with custom capabilities. For added measure, it ships with a library of pre-packaged extensions. One plugin, MCP Apps, makes it easier to manage long-running AI agent workflows, while a tool called EMA streamlines certain cybersecurity tasks.

Image: Anthropic