Wrong, you're assuming the "bespoke glue code" does not exist with MCP, it's because LLM does that for you, but it's still dealing with your logic. Actually, debugging why the LLM's requests are failing requires more work and needs to expose more details on how to handle the error. Typically requires checking the LLM's understanding of the tool definition (you're debugging the LLM's understanding, which differs across models). But also, remember about costs. Everyone will put an LLM in front of REST, with a cost per token, when a simple form could do the same? It's about unit economies, not interoperability.
