See where your Claude Code usage went

What the CLIs tell you

On every run, Claude Code reports how much of each window is used and when it resets (a rate_limit_event in its stream), and Codex writes the same into its session file. That is a total for the account. With several sessions open, you can’t tell which of them used it, or whether a session used much more than it should have.

Where it went

1. Start wisp.
   npxnpmcurlbrewDownload

   $npx wisp-agentsCopy

   macOS · Linux and Windows in beta · Node 20+ · MIT

2. Run your tasks as usual. Each time the CLI reports its windows, wisp shares the change out among the tasks that ran on that account since the last report, by how long each one ran.
3. A task’s share. Its details show it, longest window first: 22% of weekly · 31% of 5-hour. A subagent’s share counts on its lead.
4. The account’s share. Settings → Usage → Plan budget shows each Claude Code and Codex account’s windows, when each resets and when you’d run out at this pace. Under it, Where the weekly window went lists the three tasks that used the most this period (click one to open it), the rest in one row, and Outside wisp or earlier for what wisp didn’t run.

When your plan drains faster than usual

Turn on Settings → Usage → Drain watch (off by default). It compares an account’s runs of the last two hours with its runs of the two weeks before, as share of a window per unit of work: Claude Code’s reported cost, else tokens. At twice the usual rate or more, the account’s block in Plan budget gets a card:

Using your plan 3.1× faster than usual
Since 14:05, 4 runs took 18% of the weekly window for $2.10 of work. That much work usually takes 6%.

Runs that ended in an error and still used the window are counted apart. Copy report puts the numbers on the clipboard in plain words, ready for a support ticket: the harness and account, the window and its reset in UTC, the runs, the work, and what that work took over the two weeks before. Dismiss hides the card until a later run shows the drain again. With ntfy or Telegram set up, your phone hears of it too, at most once every two hours per window.

Every account’s windows on one page, no menu-bar app

With a work and a personal plan on Claude Code and a Codex plan, that’s six windows to keep an eye on, and the usual answer is a menu-bar app per vendor that polls each account. Plan budget lists them all in one place: each account’s 5-hour and weekly windows, how much is gone, when each resets and when you’d run out at this pace. It’s the same page on your desktop, in a browser and on your phone, because it’s part of wisp.

What each turn took

Every finished turn carries a receipt in its trail bar: 38k tokens · 71% cache · 3% of weekly. That’s the tokens read and written (Claude’s cache reads and writes counted in), the share read from the cache, and the turn’s share of the account’s longest window as wisp traced it.

• Past your plan. A run billed beyond the plan (Claude Code’s extra usage, Codex credits, or Claude Code using an API key instead of your plan) says Past your plan or API credit on the turn. Plan budget counts how many runs did that in the last 7 days, with Open it for the last one.
• Subagents. When Claude Code ran subagents, the trail bar says so beside the receipt: 3 web-fetch agents · Opus 5.5 · 27k of 41k output. Its tooltip lists every subagent type and the tokens per model. Where it went: subagents in Plan budget shows each subagent type’s share of the account’s longest window.
• Page reads on another model. When a run’s tools had no WebFetch but a web-fetch agent ran, the turn says Claude Code ran page reads as an agent on <model> on this account. That’s a Claude Code experiment (anthropics/claude-code#106743) that can drain a plan faster.

Two accounts, the same work

With two or more accounts on a harness, Settings → Usage shows each account’s median turn per model over the last 7 days, once each has 5 finished turns on that model: output and input tokens and the share read from the cache, with a line such as Work uses 1.9× the output tokens per turn of Personal (n=14 / 11). When Claude Code ran with different tools on the two accounts (one’s settings deny WebFetch, an MCP server only one has), the line says that too. Copy report puts both accounts’ numbers in one block for a bug report to the vendor.

Keep part of your plan for yourself

Automations stop at (Off, 50%, 75% or 90%, and a line per harness) holds back work nobody started by hand once a window passes the mark: crew check-ins, triggers, goal rounds and agents’ own follow-ups wait for the reset. Anything you start yourself still runs. wisp never switches accounts or goes past a provider’s limit on its own.