The Context Protocol Model offers welcome and powerful standardization to connect LLM to business data and tools.
Some experts talk about it like USB C of AI. Unveiled at the end of 2024 by Anthropic, the MCP, for Model Context Protocol, begins to be massively adopted in automation flows. A new universal communication protocol which promises to give all its usefulness to AI by connecting it to data sources and the various business tools. A small revolution that the JDN deciphers you.
How does MCP work?
The MCP is a new communication protocol which gives context to an AI model (LLM) and gives it tools to act on external data or on applications, all in real time. The MCP is based on the basic principle of the web: a customer-server architecture. Three elements make up the protocol. First, the host: this is the basic application that communicates with the AI model. For example Claude Desktop for access to Claude models. This host also has his MCP client who allows him to receive and issue requests to different servers. Technically, requests are sent in JSON-RPC 2.0. Finally, the central part of the system, the MCP server, which can be deployed locally or in the cloud, manages connections with the databases or the applications to which the AI model must access.
Each MCP server offers a list of tools available to the LLM to which it is connected. They allow you to consult or write data or carry out actions in the connected application. For example, an MCP server for Github can offer tools to consult the source code of a project, create outcomes or submit requests. Developers can create their own MCP servers to connect their applications to LLM. Some MCP servers are developed only on the basis of an existing API to facilitate its connection. This is particularly the case for several MCP servers dedicated to the use of Gmail, Google Agenda or Google Maps.
Concretely, once the MCP server has been implemented, Claude Desktop (to use our example) has new arrows to his bow. When the user sends him a prompt, he can choose one or more tools at his disposal. Note that in the case of Claude Desktop, each action must be validated for the first time by the user for security reasons.
How to install MCP servers with Claude Desktop?
For the time being, the most used MCP client (due to its simplicity) remains Claude Desktop, the office version of Claude. The first step is to find an MCP server that is suitable for your use. The majority of existing MCP servers are available in open source, most often on GitHub. It will then be necessary to follow the indications for the installation of the server locally on your machine. Generally, simply clone the source code (Git Clone), install dependencies and configure additional elements if necessary (example: provide a JSON API key for a Gmail MCP server).
Once the server is installed, just change the file “Claude_desktop_config.json” present in “Appdata Roaming Claude” and to enter the call configuration of the MCP server. After reviving Claude Desktop, the tools of your MCP server appear in your cat window. Claude then knows when he has to use them automatically.
MCP tools in Claude. Screenshot / JDN
In our case, for example, we have installed two MCP servers: the first “Secure-Filesystem-Server” (developed by Anthropic) allows Claude to interact directly with some files on your computer, the second “Gmail” allows Claude to have full access to our Gmail account to read, write and send emails.
Hundreds of MCP servers already available
In a few months the open source community has massively taken over the MCP protocol. Hundreds of servers have emerged for a multitude of applications and/or databases. We thus find integrations with tools as varied as SalesForce, Stripe, Spotify, notion, Discord, Telegram, databases like Mongodb, Redis, PostgreSql, but also services like Google Calendar, Gmail, Home Assistant, or even development platforms like Github, Gitlab and Azure Devops.
The areas covered are extremely wide: since project management (Linear, Monday.com), communication tools (Slack, Intercom), data analysis (Databricks, Bigquery), up to more specific cases such as the generation of images (replicate), web research (Tavily, Kagi), or even the control of domotic systems.
Even OPENAI, the main competitor of Anthropic, has just announced the upcoming adoption of the MCP in several products. SDK agents, the development kit for Openai agents, thus supports the new protocol. Agents can therefore directly communicate with all MCP servers. Finally, the Chatgpt Desktop application should soon support the protocol.
