Claude usage trackers compared: every option that works

The split that decides everything

Claude usage data exists in two unrelated places, and no amount of cleverness moves data from one to the other.

The seven categories

Start with what is already installed

Two of these ship with the product and cost nothing to try.

/usage

The community tools worth knowing

Claude Code Usage Monitor is the notable recent change in this list. Its 4.0 release moved to the official status line rate_limits as its primary data source rather than inferring limits from token counts, which is the difference between a guess and a reading.

Setting up OpenTelemetry

If you already run a metrics stack, this is the highest-leverage option, because Claude Code becomes just another service in the dashboards you already look at. It is also the only per-user route that works on Bedrock, Google Cloud, and Microsoft Foundry.

export CLAUDE_CODE_ENABLE_TELEMETRY=1
export OTEL_METRICS_EXPORTER=otlp
export OTEL_LOGS_EXPORTER=otlp
export OTEL_EXPORTER_OTLP_PROTOCOL=grpc
export OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:4317

You get eight metrics, including claude_code.token.usage, claude_code.cost.usage, claude_code.session.count, claude_code.lines_of_code.count, claude_code.commit.count, and claude_code.active_time.total, plus a set of events covering prompts, tool results, API requests, and errors. Metrics export every 60 seconds by default and logs every 5 seconds.

For an organisation

Which API you get depends on how your organisation signs in, and the two are not interchangeable.

The Claude Code Analytics API returns one UTC day of per-user records per call: sessions, lines added and removed, commits, pull requests, per-tool acceptance and rejection counts, and token and cost figures by model. Data lands with up to an hour of delay and the endpoint is free for organisations with Admin API access.

What to check before you trust one

Anything that reads your Claude Code transcripts is reading a plaintext record of every file the agent opened and every command it ran. Anthropic is explicit that these files are unencrypted and that a credential printed by a tool ends up in them. That raises the bar for what you install.

1. Does it parse locally? A tracker that uploads transcripts to compute totals is exfiltrating source code and secrets to do arithmetic that runs fine on your machine. Aggregates can travel; transcripts should not.
2. Does it deduplicate? Ask, or compare it against npx ccusage@latest. A tool that reports noticeably higher tokens for the same period is summing repeated records.
3. Does it price at event time? Rates change. Claude Sonnet 5 moves from $2 and $10 per million input and output tokens to $3 and $15 on 1 September 2026, so a tracker using current rates will restate your July on that day.
4. Does it cover the agents you run? A Claude-only number is not a spend report if half your work goes through Codex or Cursor.
5. Does it keep history? Claude Code sweeps transcripts after 30 days by default. A parser with no store of its own inherits that horizon.

Picking by what you actually need

If you only need one half, take the simpler tool. ccusage is excellent and we check our numbers against it. The case for something larger begins when you are running more than one agent, more than one account, or more than one machine.

Questions people ask

Keep reading