Claude Code usage monitors: 5 tools compared, and what each is best for

Every tool here reads what Claude Code already writes on your disk, or asks Anthropic for the same plan percentage /usage shows. None of them changes what Claude Code does. They differ in when they run, what they keep, and how many agents they read.

ccusage alternatives

ccusage prints a report when you ask for one, with nothing left running. Pick Claude-Code-Usage-Monitor to watch the five-hour window live in a terminal, with the official figure once it is your status line; CodexBar for plan limits from every provider you pay for, always in the Mac menu bar; tokscale for many coding agents in one terminal view, by model, project and day; Caprock for every session on the machine, whether or not it started it, with limits across subscriptions and cost per repository, kept past Claude Code's 30-day cleanup, in a free desktop app.

Where the numbers come from

There are three sources, and each tool picks one or more:

• The transcripts. Claude Code writes every session to ~/.claude/projects/ as JSON lines, with token counts per call. ccusage, tokscale, Claude-Code-Usage-Monitor and Caprock read them. Tokens are exact; dollars are tokens times a price table, so they are list prices.
• The status line. Claude Code passes the five-hour and seven-day windows to a status-line command on Pro and Max, as a percentage used and a reset time. Claude-Code-Usage-Monitor and Caprock can install themselves there to catch the official figure instead of estimating it.
• The provider's own endpoint. CodexBar and tokscale's Usage view ask Anthropic for the plan windows through the OAuth login Claude Code already has. tokscale says it shows that response verbatim.

A five-hour window worked out from transcripts — ccusage's blocks report, Claude-Code-Usage-Monitor when the status-line figure is stale — is an estimate. It cannot see claude.ai or Claude Desktop, which draw from the same allowance.

Which number is your bill

None of the dollar figures is, on a subscription. Pro and Max bill the plan fee; every monitor here prices your tokens at API list price, which is what the same work would have cost on an API key. On an API key, the Claude Console's usage page is the bill and the monitors are a close local view of it. What each plan includes is in Claude Code pricing; how the windows reset is in Claude usage limits.

The 30-day cleanup

Claude Code deletes transcripts after 30 days by default (cleanupPeriodDays in its settings). A tool that only reads transcripts can only show what is still on disk:

• ccusage and tokscale read what is there. tokscale's README suggests raising cleanupPeriodDays so nothing is lost, and can restore older usage from an export you kept.
• Claude-Code-Usage-Monitor keeps an opt-in local warehouse with --warehouse: records by project, model and day, 365 days by default.
• Caprock stores what it reads in its own database, so a session stays on its screens after Claude Code has removed the file, and the text of it stays searchable.

More on where the files live is in Claude Code history.

Privacy

All five run on your machine and read local files. None sends your usage anywhere unless you ask, but some reach the network for other things: ccusage and tokscale fetch LiteLLM's price table, CodexBar and tokscale read quota data from Anthropic, and CodexBar polls provider status pages. tokscale's submit (or its scheduled autosubmit) sends your totals to its public leaderboard, and CodexBar reads browser cookies for web-only providers once you enable them. Caprock makes no outbound call unless you turn one on; its database holds your prompts and tool output, so treat it like your shell history.

What none of them show

• An absolute token ceiling. Claude Code reports a percentage used, not a limit in tokens. A monitor that shows "x of y tokens" for Pro or Max has estimated y; Claude-Code-Usage-Monitor labels which of its figures are official and which are estimates.
• Other machines. Each tool sees the computer it runs on. A laptop and a desktop are two separate records.
• A team's collector. All five read one machine. Claude Code can also export metrics to an OpenTelemetry backend you run, per user and, opted in, per repository: Claude Code OpenTelemetry.
• Cursor or Copilot in Caprock. Caprock reads Claude Code, Codex, OpenCode, DeepSeek Harness and Hermes Agent, and Gemini CLI sessions it starts. ccusage, CodexBar and tokscale cover far more agents and providers.

If you only need today's figure inside a session, you do not need any of them: how to check Claude Code usage covers /usage, the Console and the status line.