The Claude Code usage tracker

• Claude Code and Codex: the 5-hour and 7-day rate limit windows, each with its reset or refill countdown.
• Cursor: the monthly window with its refill date, read from your Cursor account using the sign-in already on your Mac, plus a choice of measure (Weighted, Spend, Auto, or API).
• Kimi: monthly token totals from local session logs, plus token volume in the by-day chart. Kimi exposes no local cost or quota percentage.
• Kilo Code: monthly token and cost totals from its local session database, plus token volume in the by-day chart. Kilo exposes no quota percentage.
• Qwen Code: monthly token totals and daily history from local JSONL session records. Qwen exposes no local rate-limit quota through those records, so no percentage or reset window is invented.
• Goose: calendar-month raw token totals, daily history, and recorded cost from its local SQLite sessions and usage ledger. Goose can label a stored cost provider-reported or estimated. It exposes no provider quota, so no percentage or reset window is invented.
• Cline: calendar-month persisted input-plus-output tokens and exact recorded cost from its local SQLite index and session artifacts. Daily history keeps input, output, cache read, and cache write fields separate; cache portions are already included in input and are never added twice. Cline exposes no quota percentage, limit, or reset window.
• Mistral Vibe: cumulative prompt-plus-completion snapshots from local main and nested-agent session records. Vibe stores no per-call timestamps, so calendar-month and daily views attribute each snapshot to its latest saved end time. Vibe stores no cache counters and exposes no local cost, quota percentage, or reset window through those records.
• ZCode: calendar-month and daily tokens from exact local model_usage rows. Input includes input, cache-creation, and cache-read tokens; computed_total_tokens remains authoritative; and turn totals are not added twice. The store exposes no cost, provider quota percentage, or reset window.
• Aider: optional calendar-month and daily aggregate tokens and exact recorded message cost from an existing --analytics-log JSONL file. AgentPeek never enables or configures Aider analytics. The log has no session or working-directory attribution, so these totals are never assigned to a particular Aider history or session; cumulative total_cost remains a checkpoint and is not added as another message cost.
• MiMo Code: calendar-month and daily tokens plus recorded USD cost from persisted assistant-message counters in mimocode.db. Cache, reasoning, and request counts stay breakdowns, and no plan percentage or billing reset window is inferred.
• Kiro: calendar-month and daily tokens from Kiro's own per-turn counters. Flat session metadata takes precedence over mirrored state, and generic metering-unit summaries are not read as tokens.
• Devin: calendar-month and daily tokens from the read-only CLI session history, counted as exact response deltas with cache counters left as input subsets. Devin desktop Cascade transcripts expose no trustworthy usage, so those sessions show none.
• Command Code: calendar-month input and output tokens, cache breakdowns, daily history, and exact costUsd values from local v3 project records. Missing cost is not repriced, and the records expose no provider quota percentage or reset window.
• Hermes and OpenCode: monthly token and spend totals from their local session databases. Neither reports a quota percentage, so you get the honest count instead of a made-up bar.
• Pi: monthly token and spend totals from Pi logs. Oh My Pi: separate calendar-month and daily assistant usage from OMP journals plus parent task-result details.usage. Fork copies are deduplicated, nested artifact journals are excluded, and recorded cost is preserved.
• GitHub Copilot CLI: calendar-month tokens, request counts, and daily history from exact assistant-call rows in ~/.copilot/session-store.db. Input plus output is the total; cache and reasoning remain breakdowns. Nano-AIU credits are not USD, and the store exposes no provider quota percentage or reset window.
• Antigravity: weekly quota, reset, model, account, token, branch, and context fields from the documented CLI status-line payload. No credentials are read, and a status line you configured yourself is never overwritten.
• Grok, Factory Droid, Qoder, CodeBuddy, and Gajae Code: no usage data AgentPeek can read yet, so their sessions appear without usage numbers.

Everything is read using what's already on your Mac. No extra sign-in, no credentials, and no API keys to paste.

A reported limit window reads like a battery: green through yellow to red as you approach it. Agents whose records carry only raw token counters show the count instead, since there is no quota to fill. Turn on Count down to watch it drain from full instead. The Metrics page pairs a by-day chart of your last two weeks with a donut splitting input, output, cache write, and cache read, plus a per-model share underneath, and Refresh usage pulls the latest numbers any time. The Header and Pill settings each pick up to four agents to show in the expanded header and on the collapsed notch pill, so the numbers you care about are visible without opening anything.

Set a threshold anywhere from 50 to 100 percent and AgentPeek notifies you once when your 5-hour, 7-day, or monthly usage crosses it. That's the difference between wrapping up a task cleanly and having an agent stop mid-refactor. Daily token budgets reach every agent with daily history; daily, weekly, and monthly dollar caps reach every agent whose usage is a local token count rather than a reported quota window; and alerts for limit recovery, early limit pace, and available account handoffs round it out. Per-agent controls in Settings decide which gauges show, which account is reported when you have more than one, and how Cursor's monthly measure is computed.