See what Claude Code is doing between your prompt and its answer: every model call, every tool, each subagent on its own context window and model. Reads the logs it already writes — local, read-only. One executable, macOS/Linux/Windows, MIT.