Artificial intelligence startup Anthropic PBC today released a toolkit for connecting large language models to external systems.

The Model Context Protocol, or MCP for short, is available under an open-source license. Anthropic says the software has already been adopted by several tech firms.

Companies can connect their LLMs to external systems in a bid to make them more useful. An electronics maker, for example, could equip an LLM with the ability to answer customer support requests by giving it access to a repository of troubleshooting guides. AI models can also interact with external applications in other ways, such as by modifying the data they contain. 

Connecting an LLM to an external system usually requires writing a significant amount of custom code. Anthropic’s new MCP protocol is designed to ease the task. According to the company, it provides building blocks for integrating LLMs with external systems that spare developers the hassle of creating everything from scratch. 

Anthropic claims that MCP allows software teams to develop LLM integrations in under an hour. Claude Desktop, an application that provides access to the company’s Claude line of LLMs, can automate some of the manual work involved in the task. 

To use MCP, developers must implement it in both their LLM-powered application and the remote system that the application will access. From there, connections are established through a three-stage process. The application sends a network request to the remote MCP-enabled system, the system responds with a similar request and the application signs off on the connection with an automated acknowledgment.

MCP sends data using the JSON-RPC 2.0 protocol. The latter technology packages information into the JSON data format, which lends itself well to moving files between disparate systems.

Making data from remote systems access to an AI application is not the only use case that MCP supports. According to Anthropic, developers can use the protocol to give their LLMs access to cloud-hosted tools. A company could, for example, connect an AI programing assistant to a cloud-based development environment in which it can test the code it generates.

MCP also provides a feature called sampling. It enables an MCP-enabled server to request that an AI application perform tasks autonomously. According to Anthropic, developers can implement the feature in a way that allows the AI application’s users to review such requests before they’re processed. 

MCP already has several early adopters. Block Inc. and Apollo Inc., a startup with a popular sales platform of the same name, have implemented the protocol in some of their systems. Anthropic says that a number of venture-backed developer tooling providers are currently working on integrations of their own. 

Image: Anthropic