Thanks. My experience is something similar. When I am using claude cli - firstly, it does not use the tools all that much - and - repeated nudging does not help much. In fact, if we are using plain claude cli - there will be zero difference between having any MCP or completely raw setup.
However, I have realized things change dramatically once we change the agent. Two examples that I can readily show:
- Zed Editor. When I am using the Zed editor - and - same claude sonnet 4 model - MCP calling, overall output quality dramatically improves. Everything else same. So, what Zed is giving as input to the LLM model is quite different from what claude cli gives to LLM. And that is making a lot of difference.
- OpenCode. When I am using the opencode cli - and - same claude sonnet 4 model - suppose I am using it for fixing failed tests. The way OpenCode approaches the problem is hugely different from how the same is done via claude cli. With opencode it will fix the code in the most optimal way in the fastest time with lower tokens. Nothing certain about claude cli.
Again both these tools (Zed, OpenCode) are using LSP or some semantic tool as part of the prompt. And it is showing its impact. How can it be enhanced is the question. That is where everyone is still in exploratory mode.
I can say with reasonable conviction that these graph databases, vector embeddings etc. are not that useful - atleast for code generation.
