OpenAI is rolling out support for MCP, an open-source technology that large language models can use to perform tasks in external systems and access the data they contain.

OpenAI Chief Executive Officer Sam Altman announced the move on Wednesday. The development is notable partly because MCP was created by Anthropic PBC, the ChatGPT developer’s best-funded startup rival. 

On launch, the MCP support is available in OpenAI’s Agents SDK. It’s an open-source toolkit that developers can use to build artificial intelligence agents. Over the coming months, OpenAI will integrate MCP into ChatGPT’s desktop client and Responses API, the application programming service through which developers access its LLMs.

Companies can make their LLMs more useful by connecting them to external systems. A retailer, for example, could give a language model access to a database of product listings and use it to generate shopping advice. Building such integrations historically involved a significant amount of work.

Anthropic’s MCP protocol eases the task. It provides software building blocks that developers can use to quickly connect their LLMs to external systems. According to the company, creating integrations using the protocol takes less than an hour in some cases.

MCP enables LLMs to not only retrieve data from external systems but also perform actions in those systems. An LLM optimized for coding tasks could use the protocol to run a configuration script on a cloud instance. An AI-powered marketing tool, meanwhile, can enter ad performance metrics into an analytics application. OpenAI’s decision to add support for MCP will make such features available to ChatGPT Desktop, Responses API and Agents SDK users.

The ChatGPT developer’s Wednesday update coincided with the launch of the latest MCP release. The new version includes several feature additions.

MCP uses a technology called JSON-RPC to move data between LLMs and the systems to which they connect. According to Anthropic, the protocol’s latest release adds a feature known as JSON-RPC batching. It allows MCP to package multiple LLM data requests into one large request, which increases efficiency.

The new release also makes it easier for MCP-enabled systems to send notifications to the LLMs that access them. Additionally, Anthropic upgraded MCP’s authorization mechanism to OAuth 2.1. It’s the latest release of OAuth, a technology that helps applications establish secure connections with one another.

Against the backdrop of the updates, OpenAI investor Microsoft Corp. debuted a new MCP integration of its own. The company released a tool called Playwright MCP that combines the Anthropic-developed protocol with its own Playwright software.

Microsoft originally developed Playwright to ease the task of testing websites for bugs. The software, which is available under an open-source license, can automatically perform actions in a browser. Microsoft’s newly released Playwright MCP tool harnesses Playwright’s web browsing features to let LLMs interact with webpages.

Developers can use the tool to automate tasks such as filling forms. According to Microsoft, Playwright MCP also enables coding-optimized LLMs to automate website testing tasks. 

Image: Anthropic