herdr-fleece

herdr-fleece

Know when your herd is fleecing you. Live tokens, dollars and burn rate for every herdr pane, one screen.

English · 简体中文

Every pane in your herd, which agent it runs, whether it is working or blocked, how many tokens it has eaten, what that costs, and how fast the meter is spinning right now. The burn column is the point: cumulative cost tells you what already happened, dollars per hour tells you whether to look up from what you were doing.

npx herdr-fleece

No install, no config, no account. It reads the usage files your agents already write and asks the herdr socket which pane owns which spend.

What it shows

• One row per pane — agent kind (Claude Code, Codex, more below), herdr's own working / blocked / idle status, tokens, dollars, and a live $/hour burn rate over the trailing five minutes
• Herd totals — what the whole herd has spent and its combined burn rate, in the header where you cannot miss it
• A leak alarm — a pane that is blocked or idle while its directory keeps billing gets a red ⚠. A stuck agent that stopped costing money is patience; a stuck agent that is still spending is a runaway
• Off-herd spend — sessions running outside herdr still count. Money that vanishes from the report is the one failure a cost console must not have, so unattributed spend gets its own line instead of disappearing
• --json — the whole board as JSON, for status bars and scripts
• --all — fold in the trailing 24h of local history instead of starting from launch
• --once — render a single frame and exit

How it works

herdr-fleece joins two things your machine already has:

• Local usage files. Claude Code writes per-request token counts to ~/.claude/projects/**/*.jsonl; Codex CLI writes rollout files under ~/.codex/sessions. The parser follows the shipping conventions of ccusage's adapters, including the message-id dedupe that stops streamed rewrites from double-billing, and was cross-checked against ccusage on live data (within 2%).
• The herdr socket. One pane.list call per tick over the same NDJSON protocol every herdr tool uses. Spend is attributed to a pane by working directory, and panes claim every unclaimed session in their cwd, so a fork or subagent working in the same directory bills to the pane that spawned it.

No proxy, no wrapping your agents, no PATH games. If herdr is not running, the meter still counts local spend and says so.

Agents covered

agent	tokens	dollars	source
Claude Code	yes	yes, costUSD when present, priced otherwise	~/.claude/projects/**/*.jsonl
Codex CLI	yes	yes	$CODEX_HOME / ~/.codex/sessions/**
Gemini CLI	yes	yes	$GEMINI_DATA_DIR / ~/.gemini/tmp/*/chats/*.json
OpenCode	detected in panes	planned

Prices are a built-in table (per-model $/Mtok, cache rates included). Override any row in ~/.config/herdr-fleece/prices.json. Unknown models are counted in tokens and marked ? rather than silently priced at a guess.

Honest numbers, stated plainly

• Subscription plans. If you run Claude Code on a subscription, the dollar column is the list-price equivalent of your usage, not what you are billed. It is still the right number for "what is this herd worth per hour", it is not your invoice.
• Attribution is by directory. Two panes working in the same cwd are ambiguous by nature; the focused pane claims first. The off-herd line exists so nothing is lost either way.
• Burn rate is a trailing 5-minute window, so it reacts in seconds and cools in minutes.

Requirements

• Node 18+
• herdr for pane attribution (optional — without it you still get totals)

MIT.