# Track Claude Code usage, limits and tokens on macOS

Understand Claude Code usage tracking in Mr. Usage: session and weekly limits, reset times, local token charts, OpenCode logs and cached-token API estimates.

Source: https://usage.beffa.xyz/docs/claude-usage

## Claude plan limits in the menu bar

Mr. Usage displays Claude usage windows and reset countdowns in its Limits tab. It uses the existing Claude Code login stored in macOS Keychain, with an unexpired Anthropic OAuth login from OpenCode as a fallback, and requests the same usage information used by Claude Code’s usage command.

These percentages describe provider-reported plan allowances, not an app-generated monthly token budget. Claude Pro and Max plans have usage limits, and the amount of work a session supports varies with model, context and task. Mr. Usage observes those limits; it does not extend or bypass them.



## Claude Code and OpenCode token activity

The Claude Tokens tab reads local Claude Code transcripts under ~/.claude/projects and Anthropic messages in OpenCode’s local database. Repeated or streamed copies of a response are counted once. Charts and model breakdowns help show which recorded workloads produced the totals.

These token counts cover supported records on this Mac. They do not represent all conversations on claude.ai or Claude activity on other computers. Provider-reported plan usage can therefore differ from the locally recorded token history without either measurement being wrong.



## Why cached Claude tokens can be large

Coding agents repeatedly send context such as instructions, conversation history and tool results. Prompt caching allows repeated context to be read from cache rather than processed as entirely new input. Consequently, a workload may have a large cache-read total alongside a much smaller amount of new input or generated output.

Mr. Usage keeps uncached input, output, cache reads and cache writes separate. Claude five-minute and one-hour cache writes have different API prices. API-equivalent cost uses published model rates for those categories where a model is known; it is not your Claude subscription bill.



## Polling, throttling and saved results

Claude live requests normally start at one-minute intervals. Throttling, server errors and Retry-After responses can delay the next request. The last successful limits snapshot and cooldown persist locally across launches so restarting the app does not create a burst of requests.

When a saved result is shown, its original timestamp remains meaningful. For billing questions or unexpected provider allowances, compare the result with Claude’s own account usage page. Mr. Usage’s token charts are an additional view of supported local activity, not a replacement for the provider’s billing records.



## Related documentation

- [Mr. Usage documentation](https://usage.beffa.xyz/docs)
- [Install and set up Mr. Usage on macOS](https://usage.beffa.xyz/docs/getting-started)
- [Track Codex and OpenAI usage across devices](https://usage.beffa.xyz/docs/openai-usage)
- [AI token counts, cached input and API cost estimates](https://usage.beffa.xyz/docs/tokens-and-costs)
- [Mr. Usage privacy and optional leaderboard sharing](https://usage.beffa.xyz/docs/privacy-and-sharing)

[App source and releases](https://github.com/t1llo/mr-usage)