Today, I deep-dived into MCP (Model Context Protocol) , especially FastMCP , and
By Vishal Dhariwal · November 17, 2026 · Curated by George's Blog
Today, I deep-dived into MCP (Model Context Protocol), especially FastMCP, and came across some interesting concepts. Here are a few things I want to share.
MCP is not just about tools. It has three core primitives: Tools, Resources, and Prompts.
1. Tools (@mcp.tool())
Tools are executable functions that allow an AI agent to interact with external systems, perform operations, and take actions.
For example, an agent could use a tool to fetch customer details, create an invoice, or check the status of an order.
2. Resources (@mcp.resource())
This was particularly interesting to me.
Resources provide a standardized way to expose data and context to an AI application in a read-only manner.
For example, an enterprise agent might need access to company policies, contracts, or documentation. An MCP server can expose these as Resources, which an MCP client can discover and read when needed.
Initially, I was confused about the difference between RAG and Resources, but I realized that they solve different problems and can complement each other.
RAG is about retrieval, using techniques like semantic search and metadata filtering to find relevant information from large document collections.
For example, answering "How have the company's contract rates changed over the last three years?" might require retrieving and comparing multiple contracts. This is where RAG becomes useful.
Resources are about exposing context, not semantic retrieval. They provide a standardized interface through which MCP clients can access data using resource URIs.
In short, RAG retrieves relevant information, while MCP Resources expose data to the AI application.
3. Prompts (@mcp.prompt())
Prompts provide reusable prompt templates that can guide an AI application in performing specific tasks.
For example, we could define a reusable prompt for reviewing a contract or summarizing a financial report.
Overall, what I found interesting is that MCP provides a standardized way for AI applications to interact with tools, access context, and reuse prompts rather than building every integration from scratch.
One thing I'm still curious about is how this works in production, especially in domains like finance.
Suppose a financial firm has thousands of documents. Can we really trust OCR to extract every detail correctly? Since these details will eventually become part of our chunks and metadata, even a small extraction error could affect retrieval accuracy. Also, using an LLM for every document can be more costly.
Question : Do production systems rely on OCR, use LLMs for extraction , or follow a hybrid approach to ensure accuracy without making the process too expensive?
I'd love to hear how engineers are solving these challenges in production. Always happy to learn from the community and grow together!
#MCP #ModelContextProtocol #FastMCP #AIEngineering #GenerativeAI #RAG #LLM #AgenticAI #AIArchitecture #SoftwareEngineering