I maintain CodexMeter, an independent, open-source menu-bar companion for Codex Desktop that brings remaining quota, today's local token estimate, and the latest completed response's effective speed into separate native macOS menu-bar items.

From the start, the useful engineering question is how to make those numbers visible without making them look more authoritative than their sources allow. An account limit, a local usage counter, and a response timing estimate describe different things. CodexMeter keeps its collection paths and meanings separate.

CodexMeter menu bar with weekly quota, local tokens and estimated TPS

Original capture: weekly quota 59%, local tokens 201.2M and TPS 36. Captured example readings, not live data or a benchmark. (Source: CodexMeter maintainer)

Three signals, three different scopes

Quota percentages and reset windows come from the Codex account response. They are not calculated by converting local tokens into a remaining allowance. A countdown reaching zero also does not prove that quota has replenished: the app waits for a fresh reading. Cached or stale quota carries a marker, while unavailable data stays unavailable rather than becoming zero.

The “Today · this Mac” counter estimates input plus output tokens recorded in locally available logs during the Mac's local day. It includes identifiable Desktop sessions and subagents traceable to those sessions. Known CLI and automatic review sessions are excluded; unsupported or unidentifiable records are not guessed into the total.

Cached input is already included in input, and reasoning output is already included in output. Adding those detail counters again would double-count usage. This local estimate is neither a billing figure nor a guaranteed account-wide total across machines and cloud tasks.

Server-reported account history appears separately. Its dates retain the labels returned by Codex, and reporting can lag. Combining that history with today's local estimate would obscure their different coverage.

What effective TPS actually measures

For a completed response with defensible timing boundaries, the menu bar companion divides output tokens by measurable elapsed seconds. The result is an estimated effective response rate, not streaming decode speed or whole-task throughput.

The parser uses structural events rather than inferring timing from text length. Response identifiers and usage records help deduplicate samples. Completed tool outputs establish the next response boundary; local tool execution is excluded from that next response's time window.

Persisted timestamps still have limitations. Waiting for the first token, network delays, event buffering and remote tool activity can influence the estimate. A high number therefore cannot establish that one model is intrinsically faster than another.

Incomplete responses, unsupported event sequences and missing timing boundaries produce no new sample. Desktop samples are preferred; unattributed fallback samples carry a warning because they may include CLI activity. The recent-response chart is a local observation aid, not a model leaderboard.

Native controls without a dashboard detour

The menu bar companion uses Swift, SwiftUI and AppKit, without Electron or third-party runtime dependencies. Each metric has an independently visible native status item. Hide unused items or hold Command to drag them into position.

Clicking any item opens the shared detail panel and refreshes the readings. Accounts reporting only a weekly quota do not retain an unsupported five-hour item after a successful read. Settings include typography, colors, inline or stacked TPS, and detail-section ordering.

CodexMeter detail panel showing quota, local tokens and account history

A read-only boundary, not an offline claim

For live account information, CodexMeter starts the installed Codex App Server and exchanges JSON-RPC requests over its supported standard-input/output transport. The relevant calls include account/read, account/rateLimits/read and account/usage/read. It does not submit prompts, start coding tasks or consume reset credits.

The subprocess uses the existing Codex sign-in and makes its normal network service requests. It does not read or store authentication tokens. Calling the application wholly offline would be misleading.

The log reader follows appended bytes and preserves parsing checkpoints across restarts. Partial final records wait for a newline, and unsupported records cannot manufacture timing samples. These details matter when monitoring sessions that continue running while the app reads them.

Local parsing retains derived counters and checkpoints, not prompt or response text. It does not upload conversations, usage counters or project paths, and has no telemetry or separate cloud service. Optional GitHub update checks fetch public version metadata; automatic checks are off by default. They announce updates rather than install them.

Installation and practical trade-offs

You need macOS 14 or later and an installed, already signed-in Codex Desktop. Apple Silicon users can use the prebuilt archive or the project's custom Homebrew tap, which is separate from the official Homebrew Cask repository. Intel Macs require a source build.

The distributed application is ad-hoc signed, not Apple-notarized. Evaluate that trust decision before installation; building from source is another option. It requires no Accessibility, Screen Recording or network-interception permissions. Local quota and reset-credit reminders require macOS notification permission and appropriate refresh settings.

For broader provider coverage, CodexBar is worth examining. It also offers whole-turn output-rate and latency analytics in Usage & Spend, plus reset-credit expiry information. CodexMeter's narrower focus is a Codex Desktop companion with independently arranged per-metric items, rather than a claim that these measurements are exclusive.

The source, installation instructions, and detailed measurement limitations are in the repository.

The code is MIT licensed; OpenAI trademarks and optional logo assets are excluded from that license. The meun-bar is independent and is not an OpenAI product.