By Jordan Baillie (AI Research Scientist)
This post is the first in a series which aims to explain the ideas behind Model Context Protocol (MCP), a new AI technology which aims to make LLMs more powerful by connecting them to more relevant information and external systems.
This post and its associated white paper serve as an introduction to MCP, motivating it and discussing some of the problems it solves, as well as giving an overview of its architecture and current landscape.
Overview
The Model Context Protocol was introduced by Anthropic in November 2024 and represents a significant milestone in the evolution of artificial intelligence. As an open standard, MCP has quickly garnered widespread acclaim across the industry, with many recognising it as a truly revolutionary advancement in how AI systems interact with external tools and services. It has been described as the “moral equivalent of HTTP for the agentic web” and has spurred lots of exciting development.
The industry has responded swiftly and enthusiastically to MCP:
- Microsoft has incorporated MCP support into Copilot Studio.
- OpenAI has added it to its Responses API and is preparing to launch MCP compatibility for ChatGPT Desktop.
- Google has embedded MCP functionality into the Gemini SDK.
Many other companies have also announced plans to roll out MCP support in the near future, signalling an exciting move towards the growth and adoption of this new technology.
Which Problems Does MCP Solve?
-
MCP facilitates interoperability between multiple LLMs within multi-agent systems by standardising how different models access, update, track, and share contextual information. The diagram above illustrates that without MCP, a separate API is required to connect each LLM to every external system. However, with MCP each LLM only needs to connect to a single, unified API, the MCP server, which then handles the individual APIs for each external system.
Alongside providing pre-built connectors for popular systems such as GitHub and Google Drive, the primary objective of MCP is to address three challenges developers commonly face when working with LLMs:
- Poor Context Management: No matter how much knowledge you provide to a model, it won’t be able to use it effectively unless it has a way of retrieving the relevant parts it needs, precisely as it needs it.
- Workflow Integration: Connecting LLMs to external systems typically requires repetitive boilerplate code, increasing the risk of bugs and complicating maintenance as systems evolve and update.
- Communication Standards: Inconsistent data exchange formats between systems and models can lead to misinterpretation, debugging difficulties, and prolonged development cycles.
The Current Landscape
Ever since its inception the MCP landscape has grown significantly, with hundreds of active MCP servers listed on the official open-source MCP GitHub repository. These servers allow AI agents to do everything from accessing calendars and project management tools to creating 3D scenes in Blender.
Organizations are now building fascinating products involving MCP at a remarkable rate, reaping the substantial benefits of this new technology. For instance, the cloud-based accounting software firm Xero launched an open-source MCP server specifically designed for accounting data. This initiative facilitates tasks such as creating invoices, automating invoice management and updating client records – streamlining workflows that previously required manual intervention.
Developer tools and platforms have also adopted MCP to enhance their offerings. Companies such as Replit, Codeium and Sourcegraph have integrated MCP to provide AI agents with access to relevant coding information, improving the quality of code suggestions and corrections significantly.
Anthropic has publicly provided reference implementations of MCP servers for common systems such as Google Drive, Slack, GitHub, PostgreSQL and web automation tools like Puppeteer. These samples are a great starting point for developers looking to begin building and deploying their own MCP servers and can even be ran directly from Claude Desktop for ease of access.
Looking Ahead
Alongside the recent release of Claude 4 on May 22nd, Anthropic announced a built-in MCP connector on their API. This allows developers to connect Claude to any MCP server without writing client code – a major breakthrough for adoption.
Prior to this breakthrough, developers had to build their own client infrastructure to integrate with MCP servers and to manage MCP connections. The Anthropic API now simplifies this process by automatically handling connection setup, tool discovery, and error management. By simply including a remote MCP server URL in the API request, developers can instantly tap into advanced third-party tools, which greatly reduces the effort required to create tool-augmented agents and increases accessibility.
As the MCP community grows and continues to build more MCP servers, developers will
discover that they can easily add capabilities to their AI systems without needing to build one-off integrations. With this new update from Anthropic, the future looks bright for MCP, which is setting up to be a foundational technology. Embracing MCP is becoming essential to navigating and shaping the future of AI, paving the way for a new era of connected intelligence ready to tackle increasingly complex challenges across every industry.
Still To Come
Although still a relatively new technology, MCP has already prompted a dramatic shift in our thinking here at FD Intelligence: we are no longer constrained by isolated AI, but rather excited to explore the possibilities with connected AI.
Stay tuned for more posts where I will dive into the more technical side of MCP and explore some sample applications!