A 5-layer Claude Code setup aimed at using fewer tokens
This setup claims to cut Claude Code token use by about 40% on average in testing. The main idea is to help Claude reuse better organized project knowledge and context, instead of repeatedly reading long explanations and code files. The first layer installs codebase-memory-mcp and turns on so the project can be remembered more efficiently.
The second layer installs the context-mode plugin and runs a health check to confirm it is connected to Claude Code. The third layer installs rtk, checks the version, and initializes it globally. The fourth layer installs caveman.
The fifth layer installs -cache-fix globally and installs a cache proxy service, but the provided setup text cuts off before the full setup prompt is shown. Each layer includes a step where Claude is asked to check the linked oject and confirm that the tool is installed and integrated correctly.
Key points
- The setup claims about 40% lower token use in Claude Code on average.
- codebase-memory-mcp is used to create codebase memory with .
- context-mode is added as a plugin and checked with a doctor command.
- rtk, caveman, and -cache-fix are added as extra layers.
- Each layer includes a step asking Claude to inspect the oject and confirm .