sahil87/tu/site

tu — agent usage bundle

tu is a command-line tool that reports AI coding-assistant token cost and usage on the machine it runs on. It reads local usage data (via a vendored ccusage) and, in multi mode, aggregates across machines through a shared git repo.

This is a usage briefing for an agent operating an installed tu: when to reach for it, what each command does, how it composes, and how to read its output.

When to use

Reach for tu to answer cost/usage questions about AI coding assistants on this machine:

• "How much did Claude Code / Codex / OpenCode / Gemini / Copilot / Kimi cost today?"
• Daily, weekly, or monthly spend, as a snapshot or as history.
• Per-machine breakdowns (multi mode) and per-user views.

Do NOT reach for tu for:

• Billing management or payment — it only reports usage, it changes nothing.
• Tools it does not track (only the six sources below).
• Per-request granularity beyond what ccusage emits — tu aggregates by day.

Capabilities map

Data grammar: tu [source] [period] [display]

• source — cc (Claude Code), codex/co (Codex), oc (OpenCode),

gemini/gem (Gemini), copilot/cop (Copilot), kimi/ki (Kimi); omit for all (default).

• period — d/daily (default), w/weekly, m/monthly.
• display — bare = snapshot (current period); h/history = time series;

lb = leaderboard (users ranked by cost/tokens, multi mode); lbh = leaderboard history (period rows x user columns, multi mode).

• combined shorthands — dh (daily history), wh (weekly history),

mh (monthly history).

Examples: tu (today, all), tu cc (today, Claude Code), tu h (daily history pivot), tu cc mh (Claude Code monthly history), tu wh (weekly history), tu m lb (this month's leaderboard), tu lbh (daily leaderboard history).

Non-data commands

• tu init-conf — scaffold ~/.config/tu/tu.conf.
• tu init-metrics [url] — clone the metrics repo (multi mode); with url, also

writes metrics_repo into ~/.config/tu/tu.conf first.

• tu sync — push/pull metrics manually (multi mode).
• tu status — show config and sync state (reports single vs multi mode).
• tu update — update tu via Homebrew.
• tu shell-init <bash|zsh|fish> — emit a shell completion script for eval.
• tu help (or -h/--help) — full help text.
• tu help-dump — emit a machine-readable help contract (JSON) for tooling.
• tu skill — print this bundle (agent usage briefing) to stdout.

Composition patterns

• Shells out to `ccusage` (vendored) to read local usage data — no separate

install needed; a single ccusage binary serves all six sources.

• Shells out to `git` in multi mode to sync per-machine metrics through a

shared repo, so cost aggregates across machines.

• Shells out to `brew` for tu update (Homebrew-installed builds only).
• Is shelled out to by shll.ai's pull cron, which runs tu help-dump

against the installed binary to render tu's command reference on shll.ai.

• `tu skill` (this bundle) also renders at /tu/skill on shll.ai and is the

surface a future shll agent-setup will aggregate into agent context.

Output & exit-code contract

• stdout is data; stderr is diagnostics. Parse stdout; treat stderr as

warnings/errors. Success writes results to stdout with nothing on stderr.

• Exit codes: 0 on success; 1 on operational failure (network, git,

Homebrew, missing/misconfigured metrics repo — retry or fix the environment); 2 on usage error (bad grammar, unknown tool/shell, incompatible flags, bad flag values — fix the command line).

• Graceful degradation: when a data source is unavailable, tu warns on

stderr and falls back to the best available data (cached, local-only, or zero) rather than crashing — so a non-empty stderr can accompany exit 0.

• Machine-readable formats (data commands only, mutually exclusive):
• --json / -j — JSON.
• --csv — CSV.
• --md — Markdown table.

Prefer --json for programmatic parsing. Bare output is a formatted table.

Gotchas

• Cached data. Fetches are cached (~60s TTL). Pass --fresh / -f to bypass

the cache and refetch.

• Single vs multi mode. Behavior depends on ~/.config/tu/tu.conf (legacy

~/.tu.conf is read as a fallback with a stderr deprecation warning; an optional ~/.config/tu/org.conf supplies org-wide defaults under it): single mode reads only local data; multi mode aggregates across machines via the metrics repo. Run tu status to see which mode is active. --user / -u (including -u all, which sums every user in the metrics repo from synced data) and multi-machine views apply in multi mode only (warned-and-ignored in single mode).

• `--watch` / `-w` is an interactive TUI (live-refreshing display). Do not

invoke it from an agent — it does not terminate on its own and produces no parseable single-shot output.

• Use `--no-color` for clean, parseable output (disables ANSI color codes).
• Implicit 3-month cap. Daily and weekly history default to the last ~3

calendar months. Pass --full for the complete history. Monthly history (mh) is never capped.

• History-only flags. --since / -s and --until bound a history window

(YYYY-MM-DD or YYYYMMDD); on a snapshot display they warn and are ignored. On lb they replace the period window silently (the Δ column then compares against the equal-length preceding window).

• Token mode. --metric tokens / -t renders history and pivot tables in

tokens instead of cost (all displays); the snapshot table keeps its Cost column in dollars — only the watch delta indicator follows the metric.

• Leaderboard (`lb` / `lbh`). Multi mode only (exits 1 in single mode) —

reads the metrics repo, so today lags until tu sync (the table footer says so). --top <n> keeps the top N rows (lb, rest collapse into … +k others) or the top N user columns (lbh, rest fold into one others column); -u <name> pins/highlights a user rather than filtering; -u all is a no-op; --by-machine ranks user/machine pairs (lb only — on lbh it warns and is ignored).