I noticed that Claude completely changes its character when it starts using MCP. Can’t have a conversation, instructions in my memory are ignored including code styles.
And it just doesn’t know how to stop. When I tell it to run tests, it will continue running and fixing them until they pass. Which is what you want often, but it can also uncontrollably spiral into its own hallucinations.
https://github.com/ezyang/codemcp Is really nice because it’s a lot cheaper, but this kind of stuff and the lack of caching is pushing me towards multiple chained agents calling the API directly.
Since Claude Code is a Terminal utility, and it can call other utilities including itself, it's possible to have one Claude prompt that orchestrates multiple other Claude prompts.
So the main prompt can have a global CLAUDE.md that understands high-level tasks, while the other prompts understand better how to write, test, store, test or run code.
e.g.:
High level prompt that has a general CLAUDE.md plus description of a feature and it's edge cases as input. You discuss the design, then you execute the design, by passing to a specific code writing prompt that implements and writes tests. Then a second prompt will start running the tests, and will call another coding prompt to fix stuff if needed.
Finally the main prompt tells you either it's stuck, or the feature works as expected.
And it just doesn’t know how to stop. When I tell it to run tests, it will continue running and fixing them until they pass. Which is what you want often, but it can also uncontrollably spiral into its own hallucinations.
https://github.com/ezyang/codemcp Is really nice because it’s a lot cheaper, but this kind of stuff and the lack of caching is pushing me towards multiple chained agents calling the API directly.