MCP Deep Dive: AI's Universal Connector (Python & Salesforce Demo)

[Music] Hello and uh welcome. I'm Ash and today we're diving deep into a groundbreaking standard that's reshaping how AI systems interact with the world. The model context protocol or MCP. uh modern AI especially large language models are incredibly powerful but they often face a critical challenge uh which is accessing and utilizing real time external information uh the context they need to be truly effective imagine an AI assistant that can read your latest project files or access your company's CRM Well, that's the problem with MCP is um your resolve. Apologies. Introduced by Anthropic in late 2024, MCP is an open standard designed to create a universal framework for AI models to seamlessly connect with external tools, data sources, and systems. Think of it as the USBC of AI apps. a standardized way for AI to plug in into the um vast ecosystem of informations and capabilities that exist outside of its own training. It's about moving beyond patchy custom integrations to a world where AI assistant can truly and efficiently access any relevant context. So why was MCP necessary? Before MCP developers faced what anthropic call the N crossm data integration problem. So for every N AI models you needed you might uh need to build custom connectors for M different data sources or tools. This is incredibly inefficient, hard to scale and creates fragile systems. MCP replaces this fragmented approach with a single standardized protocol. An AI application uh uh in this case an MCP client can connect to any MCP server exposing a tool or data source without needing bespoke code for each connections. Um the standardization is crucial. It means better reproducibility as all necessary details like data set and environment specs can be managed centrally. It fosters collaboration allowing companies to share specialized AI tools or custom data sources more easily. And importantly, it enables AI to have persistent memory and context across sessions, making interaction more coherent and personalized. You might be wondering how MCP relates to other u existing concepts like MLOps or orchestration platforms like Kubernetes. Um it's important to understand they are complimentary not competing. MLOps focuses on the life cycle of machine learning models um developments uh deployment and monitoring. Kubernetes is a container orchestration platform managing application deployment and scale. MCP on the other hand is a communication protocol. It standardizes how AI talks to external systems once deployed. Think of it this way. Kubernetes might be the factory manager ensuring all your machines which is u in this case your AI services are running. MLOps is the quality control and assembly line process. MCP is a standardized set of pipes and conveyor belts ensuring all machines get passed uh sorry get the right information um in context uh when they actually need it. MCP's architecture is designed for clarity and security typically following a client um host server pattern built on JSON RPC for communication. First um we have the MCP servers. These are the components that expose data tools or functionalities. Developers can wrap their existing data repositories, APIs, or even specific functions as MCP servers. For instance, a Salesforce CRM could have an MCP server exposing tools to fetch, lead data, or update records. Servers define resources for data retrieval, tools for action, and prompts as reusable interaction templates. Next um are the MCP uh clients. These are AI applications or agents that need to access external context or capabilities. An LLM powered chatbot wanting to answer a user query about their order status would act as an MCP client. Each client runs inside the host and establishes a session with a specific MCP server. And you you probably guessed this um coordinating these is the MCP host. The host acts as a container or coordinator for multiple client instances. It's crucial for managing the life cycle of these instances and enforcing security policies, things like permission, user authorization, and consent. The hosts essentially act as a control tower determining which clients can operate, how they authenticate, and which servers can connect to. It also oversees how the AI or LLM integration happens with each client, gathering and merging context as needed. This separation of concerns is vital. The host ensures security and policy enforcement. Client focus on AI task and servers provide standardized access to capabilities. The structure is designed for stateful interaction allowing for ongoing context to interactions. The communication itself u often uses standard input output for local resources or server send events the SSC for remote ones which messages typically in JSON RPC 2.0 O format. Beyond the core client host server architecture, MCP defines several key components and capabilities that kind of make it uh powerful. Uh the the first one context embedding and handling. Um at its heart, MCP is about providing context. This isn't just about raw raw data. It's about relevant information. While MCP isn't itself an embedding model, it facilitates the delivery of data that can be used to create contextual embedding. Contextual embedding represent words or data points based on their surrounding information, allowing models to grasp nuanced meaning. MCP server exposes resources like files or database entries and prompts like reusable templates that provide this rich contextual data to the LLM. Um the host can gather and merge this context for the U clients. Policy rules and um management. So security and control are paramount. The MCP host is responsible for policy management. This includes defining which clients can access which servers, what operations can they perform and enforcing user cont sorry user consent. This host mediated permission model is crucial for enterprise adoption, ensuring that AI agents operate within defined boundaries and comply with data governance uh policies like GDPR or SO 2. MCP enables AI agents to not just retrieve information but also to take action using tools exposed by the MCP servers. the control logic for deciding which tool use can reside within the AI agent the client often an LLM itself making a recent decision based on the users's request and available tools for example an agent might decide to call a scheduling meeting tool on the calendar MCP server the anthropic API for instance can automatically handle tool discovery and execution when connected to an MCP server. While not a primary focus of the core MCP specification in its initial release, the structured nature of MCP interaction lends itself well to telemetry. Every call to an MCP tool or resource is discrete observable event. Logging these interactions, what was requested, what parameters was used, what was the outcome is vital for debugging uh monitoring performance and ensuring accountability. This data forms the basis for AI operations where AI itself manages ID operations by analyzing telemetry. As MCP adoption grows, you can expect more standardized approaches to telemetry within the EOS. MCP itself doesn't dictate which AI model to use. However, it can be a foundational layer for building sophisticated model routing systems. Imagine a scenario where an incoming query is first analyzed for its intent, language or complexity based on this analysis uh which could be an MCP tool itself. The system could then route the query to the most appropriate LLM perhaps a smaller faster tool model for simpler task or larger more powerful uh models for complex reasoning. MCP would handle the communication with these different models or u the services that front them. This dynamic selection optimizes for cost performance and accuracy. The applications of MCP are vast. Here are few um real world scenarios. Um the first one is intelligent customer service. A customer service bot using MCP can connect to a CRM server to fetch a customer's purchase history and a knowledgebased server for production information and an order management server to check shipping status. all to provide a comprehensive contextaware answer. U an AI coding assistant integrated into an IDE by an MCP can access local files, connect to a GitHub MCP server, analyze pull requests, or query a documentation server for API references, all within the developer workflow. Um um a news aggregation app could use MCPS to route users requests to different content summarization model based on the users's profile. Um example preferred topics, reading levels or the type of news um example financial news to a specialized model and um this is what u we'll build shortly. Uh imagine leads coming into Salesforce. An MCPbased server can analyze the lead data uh like region, leads core or expressed intent and route the leads to the correct sales team, a specialized nurturing sequence or a specific AI model for the initial engage. All right. Um let's get practical. We are um going to show you how to build a mock MCP server in Python that simulates routing Salesforce lead. We'll use the MCP SDK for this. Um our goal to receive lead data, evaluate its context, route it to a mock model or Q and lock that decision. All right. Um let's get practical. We're going to build a mock MCP server in Python that simulates routing Salesforce lead. We use the MCP SDK for this. Our goal, receive lead data, evaluate its context, route it to a mock model or Q, and log the decision. Um, so first let's define our Salesforce lead uh what our Salesforce lead data might look like and models we're routing to. Um so you have the the lead um dictionary where uh you have the ID, the name, the company, the region, the score and the interest. And in in terms of like also uh defining our uh mock models or Q's um that's what we've uh defined it like in another dictionary um which is like high value um EMA cloud model u you know general purpose APAC model um and these are the different um definition that we have here uh the Mock Salesforce lead is a dictionary with uh the typical lead fields and u mock models uh represents the different destination or uh processing cues for our leads. Now let's create our MCP server and define a tool that will handle the routing logic. We'll import fast MPC from MCP.server package. We create an MCP server instance um named Salesforce um lead router. The the app MCP server tool decorator is key. It's it registers our route lead function as a capability the server offers. This tool takes a lead data as input. Inside we extract contextual fields like region, score and interest. Then we have some straightforward um if else if logic. This is our control logic for routing. In a real system, this could be a call to another LLM or a more complex rule engine. Crucially, we create a log entry and print it. This simulates telemetry recording what decision was made based what context and for which lead. This kind of logging is invaluable for observability. Finally, the tool returns a dictionary confirming the uh the the routing u outcome. To actually run this MCP server, you typically use MCP command line tools. For example, if you could save this file as um lead router uh lead_outer_mncp.py, PI you could run MCP dev lead uh the the file name in your terminal. This would start a server and often open the MCP inspector a tool for testing MCP servers. An MCP client perhaps another Python script. an AI agent or even a Salesforce flow making an API call um would then connect to the server and invoke invoke the the route lead tool passing in the lead data. Um so so uh what we achieved we built a basic MCP tool that takes Salesforce like lead data evaluates its context using simple rules applies um routing logic logs the decision for telemetry purposes and outputs a structured data. This demonstrates how MCP can serve as a standardization communication layer for such context aware routing. The Python SDK with fastncp and decorators like addncp.tool u signifies uh significantly simplifies creating these servers. The ease of development is critical because it encourages more developers to expose their tools and data by MCP fostering a richer ecosystem. The actual routing logic inside our tool could be far more complex, perhaps even another LLM call to determine the best route. But MCP handles the standardization. Um, this simple example hints at MCP's profound impact. By standardizing how AI connects to context, MCP is paving the way for several key advancement. um enhance interoperability. AI tools and data sources can work together much more seamlessly regardless of who built uh them or what underlying technology they use. This is because MCP provides uh the common language. Um accelerated innovation developers can build more sophisticated contextware AI applications faster. They no longer need to reinvent the wheel for every single interaction. um they can leverage existing MCP servers or create new ones easily. The this directly address addresses the N crossm integration problem that we discussed earlier. More um MCP is a fundamental enabler uh for what many may call the AIntic era. In this era, AI systems can autonomously discover, learn about, and use a wide array of tools and data sources to achieve complex goals far beyond simple question answering types. The the ecosystem around MCP is rapidly expanding. We're seeing SDKs in multiple programming languages like Python, TypeScript, Java, and C. and an increasing adoption by major AI players like OpenAI, Google DeepMind alongside the open-source community. Of course, there are important consider consideration as this technology matures such as robust security for MCP servers, ensuring data privacy and maintaining human oversight necessary. However, the trajectory is very clear. MCP is poised to become a foundational building block for the next generation of intelligent interconnected AI agent. Its success will be significantly propelled by the richness of this ecosystem. The more diverse and readily available MCP servers and client tools uh there are the more indispensables uh this protocol becomes. So to recap, the model context protocol is a vital open standard designed to securely and efficiently bridge the gap between AI models and the vast world of external data and tools. It directly addresses critical integration challenges, fostering greater interoperability and enabling the development of more powerful truly contextaware AI systems. from its well-defined client host server architecture to its practical implementation. As we saw with our Python example, MCP offers a standardized scalable approach to building the connected AI of the future. If you're inspired to learn more, I highly encourage you explore the official MCP documentation at the model contextprotocol.io io and check out the Python SDK and other resources on GitHub and start experimenting and see how MCP can elevate your own AI projects. And thank you very much for joining this deep type and happy building, guys.