Learning about MCP

TODO:

• Implement sampling https://modelcontextprotocol.io/docs/concepts/sampling

• Implement roots.

Learnings

• There are two transport methods:

  • Via stdio. Here, you have to provide a path to a server binary client can use.

  • Via sse. Here, you have to provide a network address for the server.

    • This transport method seems more universal to me. It allows the server to be deployed separately than the client.

• I really wanted to use fastify for the server, but I could not make it work with the existing SDK.

  • This is a great opportunity to learn how to write such server using only Node!

• Core primitives that relate to MCP are:

  • Resources – You can think of these as data (like PDFs, database records and so on) that client can use to shape the correct response to the users query.

  • Prompts – Think of those as pre-defined prompts you sometimes see under the text box in "prompt to x" flows.

  • Tools – Those allow the LLM to perform actions on the user behalf. It is the server that calls the tool and responds with the result to the client. Clients job is to then pass the results to the LLM.

  • Sampling – To be honest, I'm unsure what those are yet.

  • Roots – Those define the "boundaries" of the server. For example, if your server exposes a tool that works with the file system, you could include specific directory in the "roots" so that the tool is "bounded" only to that directory.

  • Transports – Already wrote about them above. There is the STDIO and SSE transport. As for the messages, they are formatted via JSON-RPC 2.0.

• I really like the fact that the SDK has error-handling built-in.

  • I can throw an error in the server tool callback function, and the client SDK will handle that, and respond accordingly.