MCP Core Concepts

Now let's talk what MCP offers us and what's inside. What's in the box of a Model Context Protocol. If you want to read everything that MCP has to offer, then go to the documentation under model-context-protocol.ao and here you can read everything about the standard. Server developers, for client developers, examples and so on. This will be your guide throughout this journey. Here going through all the parts which are most important. All starts with an MCP host with its client. For example, cursor or windsup have their own clients that talks with their MCP servers. And you can install as many MCP servers as you want. But for example, for tools in cursor you can have maximum of 40 tools active. And MCP offers three things. Tools, resources and prompts. These RDPs the AI can use when it connects through MCP. To start with, tools are actions that the AI can take, like commands, such as search logs, update records and so on. Which the model runs using JSON RPC 2.0. And the tool might let the AI send a message in Slack, query a database or call an API. They are model-controlling. Meaning the AI decides when to use them. So when you call them in cursor, cursor agent decides when to use which tool. Often with human approval. So you have a human in the loop. For example, you have a tool, subtract, and you just write 4 minus 2. And then the tool gets called subtract and will kind of calculate this result and will give it to your back. Then we have the resources, which are read-only data, which the AI can look at but not change. This might include documents, database records or API responses. And it helps the model answer questions or provide context, but the AI doesn't take direct actions with them. And resource access is managed by the app. And then we have prompts. I think prompts are still not supported in cursor, for example. You can go here to model-context-protocol and then to example-clients and you have here the feature-support-matrix. And here you can see which client supports which feature. For example, if we scroll here to cursor, you can see that only the tools are supported and resources, prompts, emblems and routes are not supported. The Cloudly Desk App, for example, supports resources, prompts and tools. So all three are supported, obviously. Basically prompts are predefined instructions or templates that shapes how the AI responds. Like in chat GPT, you can set a role or a tone and it will respond according to your own instructions. That's all for our toolbox and MCP keeps things simple. We just have tools, resources and prompts and tools are widely used at the moment. And I think also the most useful feature here. And we will also focus more on tools in later sessions. That's all. Let's move on to the next slide.