@emilsystems: every shell command claude code runs, the entire output goes straight into the context. a git diff that's ten thousand tokens, the whole thing dumps in. cat, pytest, ls, all of it raw. most of it claude code doesn't need verbatim. this tool sits between claude code and the shell. it compresses the output before the context sees it. same information, eighty percent fewer tokens. claude code reasons better because it's not parsing thousands of lines of boilerplate looking for the three lines that actually matter. one install. ten milliseconds of overhead. first hour, the tool said i'd saved a hundred and forty-two thousand tokens.