Show how the LLM is involved in the MCP workflow

[hgmuc]

I would appreciate, if the high-level architecture diagrams would also show the data flow (requests, response, message payloads from the context) to and from the LLM. From the weather-server example with Claude Desktop it seems to me that Claude Desktop calls the LLM to create the tool call with the correct arguments. But I can't see how the returned forecast data will be formatted in a user-friendly way. Is Claude Desktop passing the forecast data to the LLM to get a nicely formatted answer or is Claude Desktop able to show this data in a nicely formatted way (e.g. table) out of the box?

• Which component (host, server) calls the LLM and when?
• Or is it up to developer to choose if the host or the server calls the LLM when it seems to be best from the workflow perspective?
• Are there any recommendations or best practises for getting the LLM involved?

Finally, adding the LLM to high-level diagrams would also help to understand potential risks when using MCP with sensitive data. Claude Desktop obviously uses Claude LLMs via the internet. But with sensitive data one might consider using a local LLM, which would maybe called by the MCP server.

[JonasHelming]

I was about to write the exact same text, this would be really interesting!

[jimlloyd]

See also #231. The fact that the LLM is not clearly identified makes it awkward to talk about it, especially with topologies where an MCP Server is a gateway for using another LLM. There are already MCP Servers that use the OpenAI or Gemini LLM APIs. This means that the User on Host is able to communicate directly with one LLM and indirectly with multiple secondary LLMs.