This matches my experience as well.
I'm a Spacemacs user, and Claude knows the difference between Spacemacs and stock emacs. It basically wrote tools for itself to hook into the emacs LSP package and look up C++ code (in our codebase, which is, uh, challenging for FOSS language servers) from gptel. I was completely fucking unable to do this via publicly-available MCP servers in the VSCode agent mode. In general I have been able to resolve a bunch of niggling config issues that I had just been ignoring in favor of doing Actual Work.
As you said, I can evaluate the elisp Claude wrote right there in the gptel buffer, try it out, and iterate before actually pasting it into .spacemacs. It it doesn't work, Claude knows a ton of debugging tricks I can try. trace-function, advice to dump a backtrace, etc. It knows how to do everything in org-mode. Super helpful. Way better than the AI Assistant that my employer bundles with our product (which is outperformed by the Google AI Overview).
Wait, you hooked up Claude Code to gptel, or you're talking about its models Opus, etc.?
My employer has Github Copilot, so I figured out how to turn that into a dumb Claude Sonnet token pipe.
Claude Sonnet is good at being Claude Code and it turns out it's also pretty good at calling gptel tools when hooked up to gptel as a dumb token pipe.