I recently attended GOTO conference in Copenhagen (https://gotocph.com/2025, you can find the slides for some of the presentations there) and would like to share some of the interesting topics.
AI agents and Model Context Protocol
Model context protocol (MCP) is the keyword which was mentioned numerous times. It is trending term nowadays. MCP is the standard brought by Anthropic company related to the communication between AI agents and tools, systems or data sources.
MCP is the driver of the significant trend to let LLM interact with your private resources. The example scenario is that you use local AI agent (like Claude) to ask for a task like "help me reply to this email from my inbox". It works like this: Claude is creating MCP client which connects to MCP server (which should be running on your machine locally) which can read your emails. Claude sends both your request and the information about ability to read emails to remote LLM which in turn is asking Claude for providing the email message in question. Claude retrieves the mail content from MCP server and sends it to LLM which can then generate the right mail response.
Very nice explanation is here:
https://medium.com/@jamestang/deep-dive-understanding-mcp-client-server-communication-with-agent-and-llm-aa4782a65991
I must say I personally didn't try it yet. Maybe I will in the future.
I think about it as about AI in general: this is both a chance and the danger. The chance if we use it wisely: don't take the LLM output as the only source of truth. Give very specific tasks to do. Understand what you get and verify it before proceeding. However it is a danger when you forget to narrow down the context, ask for a too general task to be done or simply miss the hallucination produced. You can end up with so called workslop, useless outcome which only pretends the desired result.
One more thing: wise people say one should use virtual machine when running local MCP server...
Top comments (0)