CodexBar turns AI coding limits into a menu bar — and exposes the industry's meter problem

Peter Steinberger released CodexBar on July 5, a free macOS menu bar app that polls usage limits from over 40 AI coding providers. The app shows session, weekly, and monthly windows with reset countdowns. It tracks credits, spend, and cost scans. It surfaces provider status incidents. It does not require the user to log into anything — it reuses existing provider sessions, OAuth tokens, API keys, browser cookies, or local files.

The app is a piece of utility software. It is also an indictment.

CodexBar exists because the AI coding tool market has not solved a basic consumer problem: telling the customer what they have used and what they will pay. OpenAI Codex, Claude Code, Cursor, Copilot, Gemini, Grok, Zed, Windsurf, Augment, OpenRouter, and dozens of others each expose usage data through their own portal, their own auth flow, their own API endpoint, their own definition of a “reset window.” Some providers offer OAuth. Some require browser cookies. Some require a CLI installed. Some require a manual API key. Some require the user to scrape their own browser’s local storage.

Steinberger, a well-known iOS developer and former PSPDFKit founder, built the bridge. The fact that he had to build it at all is the story.

The pricing problem in AI coding tools is not that they are expensive. It is that the cost is opaque. A developer using Claude Code through Anthropic’s API pays per token. A developer using Claude Code through Anthropic’s consumer plan pays a flat monthly fee with a usage cap. A developer using Claude through Cursor pays a different flat fee with a different cap. A developer using Claude through OpenRouter pays a per-token markup. The same model, the same work, radically different billing structures, none of them visible from the tool itself.

CodexBar surfaces 17 different provider-specific billing mechanisms in its readme. Some track “credits.” Some track “tokens.” Some track “requests.” Some track “character credits.” Some track “voice slot usage.” Some track “edit-prediction quota.” Some track “overdue invoices.” The variety is not a feature. It is a sign that the industry has not converged on a unit of account.

OpenAI Codex and Anthropic Claude are the two most-used coding agents in the market. Both have usage limits. Both reset on schedules that are not obvious to the user. CodexBar shows a countdown to the next reset. That is a three-line feature. No provider ships it.

The app’s provider list reads like a census of the AI coding tool ecosystem. Codex, OpenAI, Claude, Cursor, Gemini, Copilot, Grok, GroqCloud, ElevenLabs, Deepgram, z.ai, MiniMax, Kiro, Zed, Vertex AI, Augment, OpenRouter, LiteLLM, LLM Proxy, Codebuff, Command Code, AWS Bedrock. Each has its own auth model. Each has its own data format. Each requires a separate integration. Steinberger built them all. The GitHub repository shows the work: the providers directory contains individual modules for each service, each handling cookie decryption, API polling, rate-limit parsing, and reset-time calculation.

The privacy design is worth noting. CodexBar is “privacy-first” by design — it reuses existing sessions rather than storing passwords. It reads known file locations for browser cookies and provider configs. It does not crawl the filesystem. The app requests macOS Full Disk Access only for Safari cookie reading, and Keychain access for Chromium cookie decryption and OAuth credential storage. Steinberger documents each permission and how to revoke it. That level of transparency is rare in commercial software and almost unheard-of in open-source utility apps.

The deeper implication is for the AI coding tool market itself. If a third-party utility can extract usage data from 40 providers, the data is not proprietary. The data is not secret. The data is merely inconvenient to access. Providers have chosen not to surface it in the tool because there is no competitive pressure to do so. A developer does not know they are approaching a limit until they hit it. That is by design.

CodexBar changes the incentive. Once a developer sees the countdown, they can plan around it. They can switch providers before the limit hits. They can compare effective costs across providers. The menu bar becomes a pricing dashboard. The pricing dashboard becomes a switching mechanism.

The app is open source under an MIT license. Steinberger accepts contributions for new providers. The provider authoring guide is in the repository. The barrier to entry for a new coding tool is now lower — but the expectation is higher. Any new provider that ships without a usage API will be the one missing from the menu bar.

CodexBar is a small app. It does one thing. It does it across 40 providers. The number of providers is the headline. The fact that it had to exist is the takeaway.