Two ways to see what your Claude Code usage costs:
/meter: a pane inside Claude Code with your spend, pace, token mix, forecast, and plan week, live as you work./meter web: a full report in your browser. It reads every Claude Code transcript on your computer, back to the oldest one Claude Code kept, and evaluates your habits: cache hits, cold restarts, context size, subagents, models, projects, the hours you work, your most expensive sessions, and the tools Claude calls most.
Screenshots use the made-up history from --demo. Run npx github:Siigari/usage-panel --demo to see it yourself.
Type /meter in a session for this pane beside the conversation:
SPEND list $ PRESSURE TODAY'S MIX FORECAST PLAN whole account
5m $1.19 · 10 req now $14.33/h reads $76.52 · 72% end of day ~$149.47 5-hour 12% ↻ 11:20 PM
30m $8.68 · 74 req 1h pace $12.15/h writes $12.97 · 12% next 24h ~$291.55 ━━━━━━━━━━━━━━━━━━━━━━━━━━━
1h $12.15 · 103 req req/h 103 input $7.39 · 7% context 479K / 1M weekly 10% ↻ Thu 11 AM
today $105.64 · 1077 req cache hit 100% output $8.76 · 8% full in ~5.7 h ━━━━━━━━━━━━━━━━━━━━━━━━━━━
forecast full Sun 6 AM
■ reads ■ writes ■ input ■ output ─ context
[stacked bars: dollars per 5 minutes, with this session's context as a line]
THIS WEEK grey = even pace · white = recorded · dashed = forecast · top = 100%
[the plan week from start to reset: even pace, what you've used, and where two paces land]
now 10%, 4 pts behind even pace
week pace → 18% at reset
3h pace → 140% at reset (full Sun 6 AM)
resets Thu 11 AM
You need Claude Code 2.1.287 or newer (claude --version; claude update if older).
claude plugin marketplace add Siigari/usage-panel
claude plugin install usage-panel@siigari
Or from inside a session: /plugin install usage-panel --marketplace Siigari/usage-panel.
Restart Claude Code (or run /reload-plugins), then type /meter to open the pane. /meter again, or Esc, closes it.
/meter web (or npx github:Siigari/usage-panel from any shell, no install) reads ~/.claude/projects/, writes ~/.claude/usage-reader/report.html, and opens it. It needs Node.js 18 or newer. --demo shows the report with a made-up history instead of yours.
The first run reads everything (about 10 seconds for 12 GB of transcripts); after that only the sessions that changed are reread. npx github:Siigari/usage-panel --watch keeps it fresh: it rescans every minute and the page reloads itself.
The report grades the last 30 days on four things and lists what stands out, most expensive first:
- Cache hit: how much of every prompt came from the cache (cheap) rather than being written fresh.
- Cold restarts: a session picked up after its cache expired (5 minutes or 1 hour, depending on the session), so its whole context was written again. The dollar figure is the extra over reading it.
- Big context: the share of spend on requests carrying over 200K tokens. Every message rereads the whole conversation, so long sessions cost more per message.
- Subagents: the share of spend in subagents, which each start their own context.
It also shows spend per day (stacked by reads, writes, input, and output), a weekday-by-hour heatmap, spend by project, model, and app, context size buckets, tool calls, and the 40 most expensive sessions.
The plan percent on the report comes from the same account-usage call Claude Code's /usage screen makes, using the login Claude Code saved on this computer (--offline skips it). The week chart's recorded line comes from readings the report and the /meter pane save. Its red 3h pace is measured from those readings once they cover 45 minutes of the last 3 hours; until then it is estimated from spend (the week's percent per list-price dollar so far, times what the last 3 hours cost) and labeled "est."
The pane draws in claude in a terminal (VS Code's built-in terminal counts) and in the Code tab of the Claude Desktop app. The VS Code extension's chat panel doesn't draw mods, but sessions there still record their usage, so the numbers stay complete.
In a wide terminal the pane sits beside the conversation; in a narrow one it sits above the prompt. Ctrl+X then an arrow key resizes it.
- SPEND: what the requests in that window would cost at API list prices, and how many there were. "today" starts at your local midnight.
- PRESSURE: "now" is the last 5 minutes scaled to an hour; "1h pace" is the last hour; "cache hit" is the share of prompt tokens served from the cache in the last hour.
- TODAY'S MIX: today's dollars split into cache reads, cache writes, fresh input, and output.
- FORECAST: today plus the last hour's pace until midnight, the next 24 hours at that pace, this session's context, and how long until it fills at the last hour's growth.
- PLAN: your account's 5-hour and weekly percent and when each resets, as Claude Code itself reads them. "forecast" is when the week runs out at the last 3 hours' pace (or the week's average pace before there are 3 hours of readings).
- THIS WEEK: grey is even pace (0% at the start of the week to 100% at reset), white is what you've actually used, and the dashed lines carry the week's average pace (amber) and the last 3 hours' pace (red) out to the reset. Lines stop where they cross 100%.
- Dollars are list prices, not a bill. On a Pro or Max plan you don't pay per token; the dollars are a way to compare. Cache writes are counted at the 1-hour rate (2x input), which is what Claude Code uses on a subscription. A model the price table doesn't know is priced like the rest of its family.
- The plan percent is the whole account, the same number
/usageshows. It only appears on a subscription, after the first reply in a session. - Spend counts sessions on this computer that have the plugin on, from when you installed it. Other computers and claude.ai chats count toward your plan percent but not toward the dollars.
- Nothing leaves your machine. The pane keeps two days of requests and eight days of plan readings in Claude Code's plugin store (
~/.claude/plugins/store/). The report is a file on your disk; its only network call is the plan-percent reading from Anthropic, which--offlineskips.
claude --plugin-dir ./usage-panel # load it from a checkout; edits reload on save
claude plugin validate --strict . # what Claude Code reads from the hooks
claude plugin test # tests in tests/ (set SHOW = true in the test to print the charts)
node reader/usage-reader.js --json # the report's data, without the page
Built and tested with Claude Code 2.1.295.