Model context protocol: Anthropic wants to bring AI chatbots out of isolation

With a new protocol called "Model Context Protocol" (MCP), Anthropic wants to create a standard with which AI chatbots can connect to external data sources. Chatbots should be able to both read data from these sources and feed data into them. As an example, Anthropic showed in a video on X the creation of an HTML page in the chatbot, which is stored directly in a repository on Github. MCP was published as an open source standard and, according to Anthropic, should be open to all AI models.

The developer of Claude wants to use the protocol to overcome the isolation of AI models, which have so far mostly only been able to interact with data systems via self-created interfaces. MCP is based on two pillars: The MCP servers provide data and manage data access. They also serve as a kind of bridge to services. Anthropic has made MCP servers available in a repository for popular services such as Slack, Google Maps, Fetch, Google Drive and GitHub. MCP clients, on the other hand, are the applications and workflows that access the servers.

A special feature of MCP is that it can be used to access both local databases and those on the internet. Initially, MCP can only be used on local servers, but this is set to change quickly.

An incentive for corporate customers

In the specific example with Github, various tasks were created for the chatbot. Whenever Claude interacted with the MCP server to create or save something externally, the user was asked in a dialog whether they allowed this interaction. Users can choose between a one-time confirmation and permission to interact during the entire chat.

The Model Context Protocol is apparently primarily aimed at making the use of AI models more attractive to companies. Examples cited include integration into development environments, connection to internal company systems and the automation of workflows. Anthropic has already secured a number of partners for the start.

MCP is now available to developers. Claude Enterprise customers can already connect their internal systems via the MCP server.

Will OpenAI and co. follow suit?

It remains unclear whether Anthropic's competitors prefer to rely on their own solutions instead of using MCP. OpenAI, for example, is developing "Work with apps". The function, which is currently available for ChatGPT on the Mac, can read screen content from apps such as the Xcode development environment, saving users from having to copy and paste to make specific requests. There are also currently no reliable statements on the performance of MCP.

(mki)