The MCP (Model Context Protocol) is one attempt, and currently the most recognised, at creating a standard for communication between technologies. It is an open standard and open source framework, developed and introduced by Anthropic in November 2024. A rival AI-centric protocol is Google's Agent-to-Agent ([[A2A]]) protocol.
MCP provides a standardised way for models to interact with existing data sources, such as databases or existing APIs.
![[Pasted image 20250709121014.png]]
The general architecture of MCP is a client-server model with information exchanged using JSON-RPC carried over Stdio, HTTP or SSE, as appropriate for the specific use case
- The **AI model** (the "host") makes requests for data or actions.
- The **MCP client** (embedded in the AI or its platform) manages communication.
- The **MCP server** acts as a bridge to external data sources (APIs, files, databases, etc.), retrieving information or performing actions as requested
![[Pasted image 20250715122956.png]]
[source](https://www.theregister.com/2025/07/12/ai_agent_protocols_mcp_a2a/)
The main objective of MCP is to provide a protocol to standardise how models and tools expose their capabilities, thus facilitating communication, hence its comparison the the USB-C standard. Note, this is not the only way of facilitating communication between models - the referenced article provides a link to a previous exploration of other ways of doing this, such as LangChain.
Source: https://www.theregister.com/2025/07/12/ai_agent_protocols_mcp_a2a/
The Register has undertaken some experimenting with setting up a MCP which is interesting reading: https://www.theregister.com/2025/04/21/mcp_guide/