
MCP enables LLMs to access fresh, private, and actionable information without hardwiring bespoke integrations into every application.
- Training cutoff: Models only “know” information up to their training cutoff date.
- No live data access: They cannot query the web or enterprise systems without an external bridge.
- No direct access to private sources: Internal documents and databases require a secure interface to expose data.
- Limited actionability: Models alone don’t perform external actions (send emails, update systems, or run scripts).
- Key components of MCP architecture
- How MCP enables the flow between LLMs, host applications, and external services
- MCP primitives: resources, tools, and prompts
- Real-world examples and practical benefits

- Customer (Host application + LLM): The user-facing app that needs data or an action performed.
- Waiter (MCP client): The messenger that formats and forwards requests using MCP.
- Kitchen (MCP server): The system with access to ingredients (external data and tools) that fulfills requests.
- Ingredients / tools (External services): Databases, file stores, web APIs, and other services.

- The host and LLM determine additional data or an action is required.
- The LLM issues an MCP-formatted request to the MCP client.
- The MCP client routes the request to the correct MCP server.
- The server communicates with the external resource, retrieves data or performs the action, and returns the result.
- The client relays the result back to the host application and the LLM.

| Primitive | Purpose | Common examples |
|---|---|---|
| Resources | Readable context data the LLM can query | File contents, database rows, API responses |
| Tools | Actions the LLM can request the server to perform | Send email, create task, run a script, post to a service |
| Prompts | Standardized templates or interaction patterns | Predefined request templates, structured input forms |

- File system access — securely read and write files on behalf of the model, with audit logs.
- GitHub integration — summarize commits, create pull requests, or comment on issues programmatically.
- Web search augmentation — incorporate recent web results that are outside the model’s training data.
- Database queries — fetch private customer data to provide contextually accurate answers.

| Benefit | What it enables |
|---|---|
| Standardization | Reusable integrations, less duplicated effort |
| Separation of concerns | Host and LLM focus on intent; servers handle integrations |
| Extensibility | Add new servers for more services without changing the client |
| Reduced hallucinations | Verify facts by pulling authoritative data before response generation |
| Enhanced capabilities | Combine LLM reasoning with real-world actions and live data |

MCP is an open, standard approach for connecting models to external tools and data. In this lesson/article we’ll explore how to design and implement MCP clients and servers and how to use these primitives effectively.
- Anthropic — MCP origin and resources
- Claude by Anthropic
- Example specs and community projects: search for “Model Context Protocol” or check relevant GitHub repositories for MCP implementations and examples.