I just watched @theo new video on why MCP sucks. So I thought I would explain why MCP actually doesn't suck. Protocols only exist to connect two distinct systems in a standard way. An LLM can only perform actions through tool calling. Tool calling is the standard interface. If I'm writing an application that uses an LLM and wishes to do tool calling, I don't need MCP. The code that calls the LLM can just implement the tool calling interface and be done with it. This is essentially theos use case and probably why he doesnt articulate MCP well. If you want the tool implementation to be independent of the code calling an LLM this is where a protocol comes in. Prior to MCP most approaches to independently distribute tools focus on language frameworks like langchain. No language independent way existed such that a LLM client in python could connect to a say a Java tool. But furthermore no standard way existed such that one could expose a tool publicly in a secure multi-tenant fashion. MCP solves this problem. And despite what @theo stated, MCP does define auth, it's one of the core features of the spec (yay OAuth 2.1, DCR,and DCIM 😭) If you use MCP for 1 tool or 1000 tools, it still has value. The issue most people complain about today is that there are too many tools or poorly defined tools. This is an issue independent of MCP and where a lot of exciting work like code mode is being done to address it. But there's other progressive discovery methods that exist, such as the implementations of ChatGPT apps, which is based on MCP, or Claude skills. Code mode, or progressive discovery don't remove the need for MCP, they just provide a meta approach to deal with tool scalability issues but can still be implemented with MCP. One could possibly make the argument that we just don't need to ship tools as systems of their own, but people are doing it and getting value out of it, so 🤷‍♂️