[hn8726]

This seems like a good thread to post — can someone give an intuitive explanation of how MCP is supposed to work? I tried to set it up with LM Studio and a local LLM but couldn't for the life of me figure out what I actually need to do. MCP servers seems straightforward, I guess I need an MCP client somewhere, but I'm lost on how the LLM actually knows what tools are available and how/when a translation between "get x from external tool y" translates to actually getting `x` and how the LLM can then use it.

From what I could find online, it just works in Claude desktop app and there are some online efforts for mcp clients, but even ollama maintainers are confused about the implementation (https://github.com/ollama/ollama/issues/7865)

[thebigspacefuck]

MCP really shines with a tool like Cline, Roo Code, or Cursor’s agent mode where an agent is writing code for you and needs access to tools. Some I’ve used successfully are Sentry for fixing errors or Figma for implementing a design. Most MCP servers are something you set up and run locally. It lets you set up auth once across tasks and configure which actions/tools are auto-approved so your agent can have higher autonomy during tasks. If you haven’t used Cursor or Cline you should give them a try.

[hn8726]

Thank you for hints but the answer is about the same as I can find online - it just works with existing tools like Claude app or Cursor or Cline, but I specifically want to understand the mechanisms used and how I can take advantage of MCP servers using _local_ llms.

[yunwal]

What’s wrong with the official MCP website? It guides you through building a server and an LLM-powered client. You can just have your local LLM operate the client and you’re set.

https://modelcontextprotocol.io/quickstart/server

https://modelcontextprotocol.io/quickstart/client