MCP or Model Context Protocol released by Anthropic is an open standard that's poised to revolutionize how AI systems get access to data and tools, enabling more capable, reliable, and interconnected AI agents
LLMs are incredibly powerful but inherently limited by the context they have access to. To make them truly useful, they need a way to access internal Knowledge Bases and protected resources securely and efficiently.
One method to augment this context is Retrieval-Augmented Generation or RAG. RAG attempts to retrieve relevant information from external sources and incorporate it into the LLM's response. However, RAG can be slow, yield inconsistent results, and is hard to standardize across systems.
Tool Use is another approach that allows LLMs to interact with the external world, by executing commands and interfacing with internal systems to provide more useful and contextually relevant responses to users. Tool Use has been supported by all the latest models, but each Tool needs to be implemented and customised for every model to use.
Enter MCP which is an open protocol designed to address these challenges by enabling universal data access in a standardised way. Instead of every company building its own knowledge retrieval and tool interfaces, MCP provides a common framework for LLMs to interact with both external and internal data sources.
With MCP, organizations can standardize their AI access layers, reducing redundancy and ensuring consistency. The adoption has been extraordinary with over 1500+ MCP servers already built by the community.
MCP in a Nutshell
At its core, MCP is a protocol over HTTP using JSON-RPC 2.0 allowing structured communication between LLM clients and MCP servers.
The basic architecture consists of:
- MCP Clients – LLMs or AI agents that initiate connections
- MCP Servers – Services providing requested context or performing actions.
- Transport Protocol – Stdio or HTTP with Server-Sent Events (SSE - we will focus on SSE)
MCP servers provide access to Resources, Tools and Prompts - with Resources and Tools having the most prominent use:
- Resources - allow servers to expose data, such as file contents, API responses, Images, Log files, etc. The clients can discover resources with
resources/listand read them usingresources/read - Tools - this is where most of the power comes in, being able to discover all available tools using
tools/list, and invoke them usingtools/call- supporting dynamic operations and interacttion with external systems
MCP unleashes power of AI Agents in Enterprises
Most discussion around MCP focus on personal LLM Agents, but the protocol has enormous potential for Enterprise Agentic Applications.
Agents require real-time access to structured and unstructured data sources, and MCP enables them to:
- Leverage Agentic RAG to access real-world data in a standardised manner, instead of ad-hoc retrieval methods
- Have access to thousands of tools in a standardised way
- Discover tools dynamically allowing AI agents to locate and connect to relevant MCP servers automatically (the specifics are still in discussion)
Enterprises can leverage MCP to build robust and capable agents, for example:
- Company Knowledge-Base Agent using PowerBI tool to retrieve and answer questions related to Company Key Performance Metrics (KPIs)
- A Customer Support Agent providing contextual answers to User queries based on internal information specific to the User, and use Tools to perform actions on behalf of the user - for e.g. cancelling an e-commerce order
- A Product Assistant for your product that guides the user by showcasing the product usage in an interactive manner. For example, for a Photo Editing software the Agent can perform actions based on user questions and interactions
Challenges and Open Questions in MCP
While MCP is a promising development, several challenges still remain to be tackled for enterprise adoption:
Authentication and Authorization
Enabling access to internal resources require strong AuthN and AuthZ capabilities - An Oauth2 based approach has been recently drafted in the specification, but its real-world implementation is still evolving.
Hosting and Scaling
Most MCP servers are currently stdin/stdout-based instead of HTTP based which are useful for local apps but not for enterprises. Cloud-hosted MCP servers are needed for large-scale Agentic applications.
The HTTP MCP protocol is stateful - the client maintains a 1:1 connection with the server - this enables the Agent to dynamically update resources or tools when they get added or removed in real time.
However, this limits deployment options, specifically for serverless hosting environments. In addition Stateful connections make it harder to scale.
A stateless approach makes hosting and scaling easier
Streaming Media Support
There is no support in the protocol for streaming media, limiting Agent integration with applications that need real-time Video and Audio. The MCP protocol focuses mainly on text - while recently support has been added for audio its not for streaming usecases.
A natural human interface is crucial for agentic applications, making rich and real-time media support critical for AI systems.
Observability and Monitoring
Debugging and monitoring MCP interactions will be vital for real-world deployments. Anthropic provides an Inspector tool for troubleshooting MCP servers, mainly targeting MCP server developers.
As MCP adoption grows, MCP-aware observability will be necessary to ensure performance and reliability of the Agent applications. I expect them to be integrated into popular LLM Observability platforms such as Langfuse and Phoneix.
The Future of MCP and Autonomous AI Agents
With MCP providing standardized access to knowledge and tools, AI agents are becoming more autonomous. Future developments crucial for MCPs continued growth must include:
- More robust security models for enterprise use.
- Improved streaming support for rich and real-time multimedia AI applications.
- Enhanced self-learning capabilities, where agents dynamically adapt their behaviour based on available MCP services.
MCP represents a major step forward in making AI systems more powerful, connected, and intelligent. As adoption increases, it will reshape how AI agents interact with the world, unlocking new possibilities for businesses and developers alike.
