Anthropic is trying SO hard to fix MCP...

This genuinely feels like you told Claude what the issue was and then watched it halucinate solutions.

We have officially reached the point where we are debugging these "tools" rather than doing actual work.

I'm loving the MCP videos because I feel like the general narrative is very positive about the protocol but I personally have spent many hours fighting with it. I wonder what could replace it

Imagine tools were just functions, objects, classes, APIs (the ones that already exist in your codebase) and LLMs could just use them transparently via some kind of just in time, transparent RPC that bridges the LLM’s repl and your local execution environment. No wrapping, no exposure, just giving an agent the programmatic resources it might need a la carte.

90% of my work today is to "consult", for companies who whink that the first S in SaaS stands for Slop.

As time passes, more I feel that just sticking to Cursor is the right move

These crashouts are cinema

Here's a revolutionary new idea: define the tools using Swagger

Calling tools from the shell is even better, it's literally designed for that!

You have a pretty good point there. Issue with running two separate tasks for a the same request might break and will not produce the actual correct result. This is why we are running to request for different things so that they work more accurately one at the time.