ccperf (ccperf / ccp)

Analytics TUI and OS-level reset scheduler for Claude Code. Runs in your terminal, reads your local ~/.claude/ logs, and shows token use, cost, cache efficiency, hourly breakdowns, per-project and per-model breakdowns, plus an optional cron/schtasks/launchd job that fires a claude ping on your chosen schedule so your session resets when you want.

Nothing leaves your machine. No network calls except a once-a-day price fetch (cached).

Install

pnpm install -g ccperf
# or
npm install -g ccperf

Then run ccperf (or ccp for short).

Node 20+. Works on Windows (schtasks), macOS (launchd), and Linux (cron). Falls back to a managed daemon file if none of those are available.

Quick start

ccperf                     # launch the interactive TUI
ccperf report              # one-shot text summary
ccperf report --section models
ccperf scheduler install   # install the daily ping at fire-time
ccperf scheduler test      # fire the scheduled command once, now
ccperf doctor              # check config, paths, backend availability
ccperf export --format json > usage.json
ccperf init                # re-run the setup wizard

What it shows

Fourteen tabs, covering what's actually derivable from Claude Code's local journal:

Tab	Hotkey	What it covers
Overview	1	today's API cost + tokens, burn rate, active session, live sessions
Hours	2	per-hour usage with range toggle (today / 7d / 30d / all), 30-day heatmap
Sessions	3	recent sessions with live-green indicator
Projects	4	cost and calls by project
Models	5	cost and calls by model, service tier
Cache	6	cache hit ratio (hits / hits+fresh+writes), 5m vs 1h create
Tools	7	core tools, MCP servers, top bash commands
Subagents	8	Task() spawns and their agent types
Trends	9	daily sparkline (30d), weekday, hour-of-day
Errors	e	tool errors, cap hits, telemetry failures
Latency	l	turn duration p50/p95/p99
Env	v	git branches, entrypoints, versions
Scheduler	s	install status, next fire, history log
Export	x	snapshot everything to JSON/CSV

In the Hours tab: t cycles the time range, m toggles cost ↔ calls.

Hide tabs you don't want via config.json → layout.hidden: ["trends", "env"].

Cost & cache

Cost is computed at Anthropic's pay-per-use API rates (fetched once a day, cached). On a Max subscription this is the theoretical pay-per-use value, not your actual bill.

Cache hit ratio is cache_read / (cache_read + fresh_input + cache_writes) — high values on long sessions mean the prompt cache is carrying most of the token cost.

Scheduler

Fires a cheap claude -p hello command at a fixed time each day so your session window aligns with when you actually want to work.

ccperf scheduler install

Configure in ~/.ccperf/config.json:

{
  "scheduler": {
    "resetAt": "10:00",
    "offsetMinutes": 300,
    "bufferMinutes": 0,
    "backend": "auto",
    "command": "claude -p hello",
    "workingDir": null,
    "taskName": "ccperf-scheduler"
  }
}

Backend auto-detects in this order: schtasks (Windows), launchd (macOS), cron (Linux), daemon (fallback, writes to ~/.ccperf/daemon.json).

Test it:

ccperf scheduler test            # run the command now
ccperf scheduler test --dry-run  # log it without executing
ccperf scheduler status          # see what's installed
ccperf scheduler uninstall       # remove it

Every fire is logged in ~/.ccperf/scheduler.log.

Configuration

All state lives under ~/.ccperf/ (override with CCPERF_HOME):

• config.json — scheduler, tab layout, refresh interval
• scheduler.log — fire history
• cap-hits.jsonl — detected cap hits
• daemon.json — scheduler state when using the fallback backend

On first run, ccperf init asks for:

• Your desired reset time
• The command to fire (default claude -p hello)
• Whether to install the scheduler now

Non-interactive output

ccperf report prints the same data as the TUI but as plain text, for piping or cron:

ccperf report                         # everything
ccperf report --section overview      # just one section
ccperf report --json                  # machine-readable
ccperf report --since 7d              # last 7 days only

ccperf export writes a full snapshot:

ccperf export --format json --out usage.json
ccperf export --format csv  --out usage.csv

What it's not

• Not a cloud service. Nothing is uploaded.
• Not an official Anthropic tool. Cost numbers are best-effort using published API rates.
• Not a live status line — Claude Code ships one; ccperf is the retrospective + scheduler tool you run alongside it.

Built on

cc-logs-helper — the typed parser library. If you want to build your own tool, start there.