Changelog

New

• GitHub Copilot, an eighth provider, on both Mac and iPhone. It only reports figures for a personally paid Copilot plan — GitHub exposes usage for company- or organization-managed seats at the organization level only, so those accounts show nothing on this card.
• ChatGPT numbers now come from a live fetch after a one-time sign-in, instead of only reading whatever Codex had already logged locally on this Mac.
• A balance figure on the ChatGPT card, and an outage indicator on any card whose provider is down.
• Desktop widgets, a compact menu bar capsule showing up to four values at once, a segment switcher above the cards, and a freely arranged key-figures bar.
• A tip jar under Settings → About — voluntary, unlocks nothing, changes nothing.

Fixed

• The desktop widgets froze over time. They asked to be redrawn thousands of times a day and ran into the system limit, which throttles them. This is the reason 6.4 was withdrawn before review.
• Demo mode kept its sample figures for barely a minute before the refresh timers overwrote them — and it showed some real account values instead of samples.
• A card whose sign-in had broken offered only a Retry that could never help. ChatGPT, Kimi and Copilot now point at the place where the problem can actually be fixed.
• The key-figures bar dropped a tile whenever its value was missing and rearranged itself. It now keeps the last known value in place, greyed out.
• The connections list was incomplete, in the app and on this site. Cancelled requests no longer count as a provider outage.

Released 29 August 2026. A planned 6.2 was folded into this release — its two fixes ship here.

New

• One line that answers «what does AI cost me this month, all in?» Enter your subscription prices once (Settings → Display → Subscription costs) and the cockpit adds your live API spend on top. Kimi and Grok stay out of the sum — they only report a balance, and an honest number beats a complete-looking one.
• Feedback and rating, side by side. Two equal links in Settings → About; and the app may now ask for a rating itself — rarely, only after days of successful use, never after an error.

Fixed

• The plan badge on the Claude card. It was never fetched: this client's tokens carry no plan, and the badge came solely from ~/.claude.json, the local Claude Code store. In the Mac App Store build that file is out of reach without a folder grant, and a second Claude account would have been given the first account's plan — that file knows only one. Now read from api.anthropic.com/api/oauth/profile, one field (organization.rate_limit_tier) per token, so per account, then kept in the keychain. Shipped first on iPhone as 2.2.2.
• The setup assistant appeared in German for everyone. Every one of its strings was missing from the English catalogue, so each fell back to its German key. English is this app's default language regardless of the system, which made the feature 6.1 leads with unreadable for most people meeting it. The iPhone edition was never affected — its catalogue is extracted at build time.
• Ten more catalogue gaps closed in both languages — among them the second-account description and the demo-mode section — and a new test now checks the source code against both catalogues, so this class of bug cannot slip through again.

New

• A second Claude account. Settings › Accounts now offers a second one, with its own card, its own mark ("C2") and its own colour, and both accounts can be named. Signing in a second time used to silently overwrite the first.
• A demo mode. Settings › Accounts › Try it out fills every card with sample figures — no sign-in, no keys. The iPhone edition has had one since 1.0; the Mac never did. Nothing is fetched and nothing overwritten while it runs.
• A setup assistant on first launch. Four steps walk you through what the app does, which services you use and the credentials for them. Skip works on every step.
• "Reset everything". Clears both Claude sign-ins and all API keys from the keychain, resets every setting, signs out of launch-at-login and the hotkey, deletes the recorded history and releases granted folders. Keys stay valid at the services themselves.

Changed

• The Anthropic API card has its own colour (an A mark instead of sharing Claude's orange C).
• Claude is a shade lighter in dark appearance, to leave contrast room for the second account's card.
• The Grok card no longer looks disabled — it had shared its lightness with muted secondary text.

Missing from the source changelog file, though it shipped and is live in the Mac App Store — summarised here from the site's own release notes.

• Two more providers: OpenRouter, where one account reaches hundreds of models, and Grok from xAI. Both show what you have spent and what is left.
• Cards can be dragged where you want them, within a column or across to the other one; the arrangement is remembered, with a reset in Settings.
• The menu bar now shows what you pick — two settings choose the first and second figure from every window the app currently has.
• The Anthropic API cost card was broken and is fixed: it asked the server for the wrong address and blamed your account for the empty answer. A rejected key now says plainly what it needs.
• The light theme is properly readable: cards sit lighter than the ground instead of darker.

Each card now carries its own coloured letter mark — C for Claude, G for ChatGPT, O for the OpenAI API, A for the Anthropic API, K for Kimi, S for the sessions — matching what the iPhone edition already showed.

• A cancelled fetch is no longer reported as a failure; a card keeps whatever figures it already had.
• Requests now have a real ceiling — two minutes — instead of relying on a system default that could hold a connection open for days.
• No more raw system error text on a card; wording of our own instead (no internet connection, service unreachable, secure connection failed).

• Settings moved from tabs to a sidebar, after a macOS 27 beta moved the tabs of a TabView into the title bar and collapsed whatever did not fit.
• Regression guards added for sidebar width and label wrapping.
• About now names the developer: Albert Frick.

• Settings window fixed: no longer a hard-coded 780×780 sheet, now its own resizable window with a sensible minimum.
• A second, independent cause of the same squeeze removed — the panel behind the sheet no longer resizes to the dashboard's shrinking width.
• Regression guard added; CI now runs on macOS 14, 15 and 26.

• Appearance: selectable dark (default), light or system.
• Language: selectable, English default regardless of system language.
• Mac App Store build: fully sandboxed, with security-scoped folder grants for ~/.claude and ~/.codex. Released on the Mac App Store on 18 August 2026 as 5.1 (56).
• Localization fix for ten format strings and seven core error messages that silently fell back to German.
• License: the repository is private and proprietary from here on; versions up to 5.1 remain available under GPL-3.0 to those who obtained them while the project was open source.

First public release.

• Bilingual: full German and English interface (316 localized strings).
• Renamed to AI-Cockpit, new app icon, new bundle identifier com.ipstyle.aicockpit.
• OpenAI credit balance: derived from a recorded top-up amount minus billed daily costs, chosen deliberately over storing a browser session key.
• Security review 4.2 (Claude Fable 5): no critical finding; two important ones fixed.

A planned 2.4 never reached the App Store — everything it carried ships here.

New

• GitHub Copilot, an eighth provider — the same one the Mac gets with 6.5. It only reports figures for a personally paid Copilot plan; GitHub exposes usage for company- or organization-managed seats at the organization level only.
• An outage indicator on the Claude, OpenAI and Kimi cards, read from those providers' own status pages. Two deliberate differences from the Mac: with no hover text to fall back on, the message stands as its own full line, and it is fetched with the foreground refresh passes — an iPhone runs none in the background.
• A balance figure on the ChatGPT card.
• A tip jar under Settings → About: three amounts (Espresso, Cappuccino, Dinner), repeatable, unlocking nothing and changing nothing, and the hint at the foot of the card list falls silent once you have given. With it, the download is free again — 2.3 carried a price — and the tip jar is the app's only in-app purchase. The Mac's tip jar comes with 6.5, which is still in review.

Fixed

• The widget and the watch complication froze over time. Both asked to be redrawn on every refresh tick and ran into the system limit, which throttles them — the same cause as on the Mac, with the wrist affected on top.
• A refused sign-in left a dead button. A card whose stored credentials had been rejected offered "Sign in", and pressing it did nothing. All eight cards were affected — and it was the one state that only that button leads out of.
• The Copilot card offered a Retry where the token was missing a permission. Retrying never helps there; it now points at the settings, where the token can be replaced.
• The connections list under Settings → Privacy and Security was wrong: api.github.com was missing, and api.kimi.com was listed although this edition never contacts that host.
• Five Copilot error messages were missing from the English catalogue and appeared in German on English devices.
• The Anthropic card has its own colour, and a card without a provider no longer borrows the accent colour — it was indistinguishable from the Grok card.
• The light theme had drifted away from the Mac's. Small print under the figures reached only 3.84:1 against the card, below the threshold; it is 5.43:1 now.

The first update after the debut, and a large one: 2.0, 2.1, 2.2, 2.2.1 and 2.2.2 were built but never submitted on their own — all of them ship here.

New

• OpenRouter and Grok (xAI), on the phone and on the watch. One OpenRouter account reaches hundreds of models; Grok shows the balance of the xAI account and the spending cap that applies. xAI needs a management key from the console, not the xai-… key used for model calls.
• Refresh rates per source, adjustable, with the same defaults and ranges as the Mac — and the clock now runs while you are looking, not only when the app comes to the front. Foreground only: iOS lets no app fetch in the background, and the settings page says so rather than promising an update that cannot happen.
• A setup assistant on first launch. Four steps instead of seven grey cards full of denials; Skip works on every step.
• A second Claude account with its own card, its own mark ("C2"), its own colour and a name of your choosing. Signing in a second time used to overwrite the first without a word.
• Siri and Shortcuts. "How full is my Claude window?" gets an answer in one sentence, the age of the figures included. The command never goes to the network — it reads the same store the widget draws from. A second command brings the app to the front so it can sit inside a longer shortcut.
• A Live Activity for the window that is pressing. Once the Claude card passes the warning threshold and a reset time is known, the Lock Screen and the Dynamic Island count down to it. Local only, no push token: no server is told how full anyone's subscription is. The price is named honestly — the percentage moves when the app has run, not in between.
• The Mac's monthly-total line. Enter your subscription prices once and the app adds the running API costs on top. Nothing is converted, and the line says so as soon as both parts are involved. Kimi and Grok stay out of the sum — they only report a balance.
• Feedback and rating, side by side under About, with no mood question in front of them; and the app may ask for a rating itself — rarely, and never while a visible card shows an error.

Fixed

• The plan badge on the Claude card was never filled. The stored token carries no plan — it never did, not even right after signing in — and the Mac only shows one because it reads ~/.claude.json, a file no iPhone has. It now comes from api.anthropic.com/api/oauth/profile, one field, asked once per account, so the second account gets its own.
• The accounts row said "signed in · 4 of 3 keys". The denominator was a literal from before OpenRouter and xAI; four further places carried the same stale number, among them the privacy page and its list of network destinations.
• The demo mode showed a projection the app does not have. It is out until the history that would make it real exists.
• The Grok card looked disabled — its colour sat at the same lightness as muted secondary text. And the card header wrapped at Kimi and Grok.

Changed

• The Anthropic API card has its own colour and an A mark; with a second Claude account there would otherwise have been three warm cards in a row. Claude is a shade lighter in dark appearance to make room for it.

The first release. 1.0 and 1.0.1 were submitted and replaced before they reached anyone, so what arrived is 1.1 — with the Apple Watch already in it.

New

• Five cards — Claude, ChatGPT/Codex, the OpenAI API, the Anthropic API and Kimi. Every card fetches its figures on the device itself; no Mac is needed. Cards you do not use can be hidden and the rest reordered, and each source runs for itself: if one breaks off, the others are untouched and whatever a card already had stays standing.
• Widgets in five sizes — small, medium and large for the Home Screen, circular and rectangular for the Lock Screen. One line per service in its own colour and with its own mark, label and figure kept apart, and the age of the numbers alongside.
• Apple Watch. Three levels: a verdict in one sentence, the sources as a list below it, and a ring per window in the detail — paged with the crown rather than scrolled, because scrolling cut the second ring through its own number. Plus a complication in all four shapes, legible in the single-colour rendering because every warning level has its own symbol, not just its own colour.
• The watch shows, it does not fetch. No token, no key and no secret lives on it; finished display values arrive from the iPhone over Apple's own pairing. A refresh token may be redeemed in one place only — a third party in the race would eventually lose and sign you out.
• No empty start. App and widget open with the last known figures and write down how old they are, instead of going blank while fetching.
• A demo mode with a full set of plausible figures and no sign-in at all.
• Notifications at reached limits and starting windows, each one switchable and all of them off to begin with.

Fixed

• Pull to refresh cancelled the fetches it had started. Every source ran in one group whose root was SwiftUI's task, so when SwiftUI discarded it they broke off together and three cards read "cancelled" although no service was down. Each source now runs in a task the app owns.
• The ChatGPT sign-in hung silently: the callback path differs per service (Claude /callback, Codex /auth/callback) and was hard-wired.
• Requests had no real ceiling. timeoutInterval only measures the pause between two signs of life, so a cost fetch could run for 35 minutes and hold its source as "in flight".
• Hidden cards and newly entered keys took effect only after a restart — a global fetch lock swallowed the second pass. Each source now locks for itself.