Agamemnon reads local wire logs, SQLite caches, and JSONL transcripts from every coding CLI it can find. Cost is priced per model with the real cache multipliers, so a cache-heavy day reads true instead of triple. Windows are measured against the limits the CLIs actually report, never guessed, and labelled measured, estimated, or user-set. A suggestions tab flags cache thrash, wasteful model mixes, and limit pacing, all computed locally. Everything stays in ~/Library/Application Support/Agamemnon/.
full cache split, plus real limit hits and reset times
wire logs, per-turn usage records
rollout transcripts, per-turn delta
per-message token counts
usage nested in the message store
session-level totals
output tokens only, and the card says so
no local token counts, so requests and lines instead of a fabricated zero
git clone https://github.com/anishfyi/agamemnon.git cd agamemnon ./build.sh open Agamemnon.app