When does my Claude Code quota reset?

The provider decides. Agent HUD shows the reset time that arrived with the window and only marks the window reset once a new reading confirms it, either by moving the reset time forward or by showing the window full again. Separate windows on one account reset on their own schedules.

Does the percentage mean used or remaining?

Used. A window at 90% has 10% of its allowance left. This follows the convention of Claude Code's /usage.

Why does one of my windows show a dash instead of a number?

No reading has arrived for it yet, or the provider does not expose that figure for your account. A dash is never the same as zero, and Agent HUD will not estimate the number from token counts.

Do Codex Desktop and the Codex CLI use separate quotas?

No. They sign in to the same account, so their windows are one set of rows and usage from one is never duplicated into the other.

Can I change the 70% and 90% alert levels?

No. The thresholds are fixed and are not stored or synchronized, so every device shows the same state for the same reading.

Agent HUD · Guides

Quota windows and resets

Every provider decides what your limits are and when they refresh. Agent HUD shows what the service reports, and nothing it did not say.

A window is whatever the service reports

A quota window is one row per window the provider returns, each with its own reset time and its own period. Agent HUD does not decide what your limits are, and it does not fill the rows from a template: a window the service calls primary is not assumed to mean five hours, and a window named by its period — minutes, 7d, weekly, monthly, MCP — keeps the name the service gave it.

One account's windows appear once even when two programs share that account. Codex Desktop and the Codex CLI sign in to the same place, so their usage is one set of rows rather than two, and usage recorded by one client is never copied into another client's account.

Every percentage is the share you have used

This follows the convention of Claude Code's /usage. Providers usually report what remains, and the app converts, so a window at 82% has 18% of its allowance left.

A reading that has not arrived is shown as an em dash, never as 0. Zero appears only when the service actually reported zero.

A reset is only real once a new reading confirms it

When a window's reset time passes, the row keeps its last reading and the moment that reading was observed. The deadline is treated as pending, not as done.

A reset counts as confirmed when a new reading either moves the reset time forward or shows the window full again. Until one of those arrives, nothing claims your quota came back.

This matters because a deadline is not an event. Provider-side timing, clock differences and a refresh that has not run yet all mean the moment on the clock can pass before the allowance actually returns.

What is never inferred

Tokens are never converted into quota, and an unavailable quota is never derived from token counts.
Claude Code is the one client that can show a share for the current session: its share of the tokens in the running five-hour window, times that window's utilization. Every other client shows an em dash.
One API response counts once whatever the log layout. Copies of the same event from two files, or from two Macs, merge into one; two genuinely separate requests that happen to carry identical counts are both kept.

Reset credits and balances are not windows

Codex reports how many resets are available, and that number is shown exactly as the service gives it. The per-credit expiry list beside it is supplementary and never used to work the count out.

Clients billed per request, such as DeepSeek, have no window at all. They show an account balance and an estimated cost for recent usage, and an estimate is always labelled as one. Money keeps its currency: it is never converted between currencies and never turned into a percentage.

When Agent HUD warns you

Warning starts at 70% used and critical at 90%. The policy is fixed — nothing to configure, nothing stored, nothing synced between devices.

A colour describes the state of one reading. Readings from different windows are never combined into a single health score, and a colour never says anything about how a task is going.

Four things each notify once: crossing 90% used, reaching zero, a forecast that the window will run out before its reset, and a confirmed reset.
The first reading of a window after launch is a silent baseline, so whatever is already on screen at start-up never raises an alert.
Readings older than 30 minutes stay visible but stop generating alerts.

Common questions

When does my Claude Code quota reset?: The provider decides. Agent HUD shows the reset time that arrived with the window and only marks the window reset once a new reading confirms it, either by moving the reset time forward or by showing the window full again. Separate windows on one account reset on their own schedules.
Does the percentage mean used or remaining?: Used. A window at 90% has 10% of its allowance left. This follows the convention of Claude Code's /usage.
Why does one of my windows show a dash instead of a number?: No reading has arrived for it yet, or the provider does not expose that figure for your account. A dash is never the same as zero, and Agent HUD will not estimate the number from token counts.
Do Codex Desktop and the Codex CLI use separate quotas?: No. They sign in to the same account, so their windows are one set of rows and usage from one is never duplicated into the other.
Can I change the 70% and 90% alert levels?: No. The thresholds are fixed and are not stored or synchronized, so every device shows the same state for the same reading.

Guides

Last updated: 2026-09-20 · The behaviour described here is documented in full, with the code it comes from, in Agent HUD Open. github.com/jazzenchen/agent-hud-open