What's new in AI Cockpit for Mac 7.7

Submitted to Apple on 18 October 2026 — a free update once it is released. Until then 7.6 is the current version; it and the versions before it are further down this page.

One card for all your Claude Code and Codex sessions, Claude figures straight from Claude Code, OpenAI costs that hold up, and a Diagnostics tab that tells you why.

New in 7.7 — one sessions card for Claude Code and Codex

The sessions card is rebuilt. It shows your Claude Code and your Codex sessions together and sorts them by what you have to do next.

• Three groups. Needs you comes first, in strong red with an exclamation mark: sessions that wait for an answer or an approval. Still running follows, grouped by project. Idle comes last, in grey.
• Waiting sessions stay in view. A session that waits for you stays listed for up to 24 hours, so a night away does not hide it.
• Mark as seen. Click the tick or the block, and the session moves from Needs you to Idle. A new reply brings it back in red. Mark all as seen does this for every waiting session at once, and a right-click on an idle session marks it as “needs you” again. The counter and the one-line summary count only sessions you have not seen. The marks live in this app’s settings on your Mac and drop out after 48 hours.
• Running sessions in full. Each row shows title and project, model and effort, tokens and the context window — with a warning when it gets tight — and the share of the current 5-hour window. Subagents hang under their session as a tree: running, done after a time, failed, or “no response”. Background agents that still run are recognised, too.
• Idle sessions keep their detail. They show the same facts as running ones, only quieter: “idle for 2 h”, the last quote, a context warning and badges for subagents.
• All, Claude Code or Codex. A filter above the list switches between them. Each view shows every session in full — no cut-off, no “and 12 more”. The setting “At most n sessions” is gone.
• Codex needs its folder. The app reads Codex sessions from your .codex folder, after you allow it once — the same folder access that the ChatGPT quotas use. Without it the card says “Codex: access missing”. If the app has to leave Codex files out, it says so and names an upper bound instead of hiding it.
• Fewer buttons. “Open project folder” and “Show transcript” are no longer buttons on every row. A right-click still opens the project folder or copies the session ID.

New in 7.7 — a window you can resize

• Pull the dashboard to the size you like. Drag a corner; the app remembers the size for the next time. The window never grows past the visible screen and never gets smaller than the card needs.

New in 7.7 — header and key figures

• Processor load and memory under the title. A small line shows how busy your Mac is and how much memory is in use. It measures every 5 seconds while the window is open; you can turn it off or change the interval (2 to 30 seconds).
• Key figure tiles that are never cut off. Each tile shows account and window on two lines, in full. When the row is full, tiles wrap into a second row. You can drag them into the order you like, and choose which ones show — with a preview, and a maximum number — in Settings.

New in 7.7 — Claude figures straight from Claude Code

Sometimes Anthropic slows down the request AI Cockpit uses for the Claude card. Claude Code still knows your 5-hour and 7-day figures: it hands them to its status line with every message. From 7.7 on, AI Cockpit can take them from there.

• Taken window by window. The 5-hour and the 7-day figure come from Claude Code as soon as they are newer than the last request. For each window the newer reading wins, so the card never shows something older than it has to. Model windows still come from the request.
• You see where a figure comes from. Point at a window and it reads “from Claude Code, 1 min ago”. Diagnostics names the source of each window.
• Only numbers are kept. Percentages, times, the Claude Code version and an ID of the account as a one-way hash — no name, no e-mail, no text. Everything stays on this Mac; nothing is sent.
• Your own status line can stay. If Claude Code already has one, the app shows a ready command that keeps it and adds the capture. AI Cockpit only reads ~/.claude/settings.json for that and never writes to it.
• A tip when it helps. After the request has been slowed twice in a row, the Claude card suggests this once. You can hide the tip.
• Off until you set it up. Without the status line nothing changes. A wrong or foreign file is ignored; then the card behaves exactly as in the first 7.7.

Set it up

New in 7.7 — clear messages, safe numbers

• “Provider temporarily overloaded.” When Anthropic (HTTP 529, “overloaded”) or OpenAI (HTTP 503, “server_is_overloaded”) says it is overloaded, the card says so instead of showing a general server error. The last figures stay, with their age. The message says nothing about your quota.
• Broken numbers are not read as zero. A number in an answer that cannot be right — say 1e400 — counts as unknown. It is never read as 0% or $0 and never shown as infinity. Absurd times or amounts no longer crash the app.

Fixed in 7.7 — the OpenAI card

A customer wrote that the OpenAI card showed “no JSON object” all the time and gave no hint why. 7.7 answers that in three steps: read less at once, name what is missing, and say why.

• Costs load in sections of up to 31 days, newest first. Before, one broken or cut-off answer from OpenAI failed the whole card. Now AI Cockpit splits a section that fails, again and again, until it finds the bad day. The rest of the card stays.
• Days it cannot read are named — never counted as zero, never added to a sum. The app retries them in the background, and the cost history survives a restart.
• Nothing is quietly wrong. An amount with a gap reads “$12.40 · 2 days missing, 118 catching up”; an unknown “today” reads “Today not available”. The budget brake says “incomplete – no verdict” instead of “normal” while days of the current month are missing. The daily cost alert stays silent until “today” and the comparison are complete. The report, the month summary, the share image, the widget and the status line use the same words.
• A gap from an earlier year no longer blocks the verdict for the current month. The remaining credit reads “cannot be calculated” while days since the top-up are missing, instead of a credit that is too high.
• Cost answers with extreme numbers are read. Some answers contain number values that the system’s JSON reader rejects, and that alone broke the card. The app now sets such values aside one by one and reads the rest.
• Amounts are checked for every provider. A missing amount counts as nothing, as before. An amount that is there but unreadable — text without a number, a value out of range — now shows as an error instead of a silent zero. Unreadable token counts read as unknown. For xAI, an unreadable spending cap shows as a quiet “Spending cap unreadable” line instead of vanishing.

New: the Diagnostics tab

The Diagnostics tab in Settings is rebuilt. It answers one question: is everything fine, and if not, why?

• One sentence and a light per provider. The sentence says whether everything is fine. Each provider has a status light — green, yellow, red or grey — with its last success, its next request and any rate limit.
• A log of the last 50 requests per provider: path, period, status, duration and size. For a failure it also gives the reason, the position in the answer, the content type and the request ID. It never holds keys, tokens, names or the content of an answer. It stays on this Mac; “Delete request log” clears it, and so does “Reset everything”.
• “Test now” runs the real request of the card and shows the steps live.
• A report for support. Copy it, share it, or open it as a mail draft. The app never sends it — you read it and send it yourself. It holds no keys and no account IDs, and amounts appear only if you switch them on.
• Settings › Accounts: “Last request failed” opens the diagnostics of that provider.

Also fixed in 7.7

• Grok (xAI): “Team ID (manual)” now works. The app saved the field but never used it. A management key without the right to read the team list failed with a 403 even with the right ID. Now the ID counts.
• ⌘, opens the real Settings window. That shortcut, and the Settings command of the system, used to open an empty window named “AI-Cockpit Settings”.
• A cancelled sign-in frees its local port at once, so a new try can start right away.
• Demo mode keeps its example sessions. As soon as the window opened, it showed real Claude Code sessions. Now only the examples stay.

Changed

• Freshness per window. Key figures, rows, the widget and the status line judge each window on its own: an old value turns pale and shows its age, while the fresh one next to it stays normal.
• Cost pages are read strictly. Anthropic and OpenAI cost answers need a clear “more pages” flag, and hitting the page limit is an error — not a quietly short total.
• Header buttons are back in one row. Collapse/expand, pause, usage by project, Your AI month and share as image are buttons again, each with a tooltip. The “…” menu from 7.6 is gone.
• A provider that has never been asked shows “Not requested yet”, not a green light. Without a single success for 24 hours it turns red.
• An unreadable answer with status 200 shows as “200 · unreadable” in the warning color. An expected 404 — the optional spend limit — shows grey as “optional – not available”.
• The status line file (stand.json) marks incomplete OpenAI costs (incomplete, missingDays, openDays); an unknown “today” is left out.
• Demo mode shows only example values in the self-test.

New and moved settings

Good to know

• A gap is not a zero. If a card says “2 days missing”, the sum holds only what OpenAI delivered. The app fetches the rest later; there is nothing for you to do.
• The report is a draft. “Send to support” opens a mail draft in your mail app. AI Cockpit sends nothing itself.
• One new file per Claude account next to the status file: ~/.claude/aicockpit/claude-limits-….json. It holds percentages, times, the Claude Code version and the account hash — nothing else. The privacy page lists it.
• 7.7 adds no new address and collects no data. The request log stays on your Mac. Still free.
• Still runs from macOS 14.
• The sessions card only reads. It opens the transcripts in your .claude and .codex folders on this Mac, after you allow it. Nothing is sent anywhere.
• iPhone, iPad and Apple Watch get the OpenAI fix with 2.18, which is in review at Apple. They have no sessions card: sessions are files on the computer, and the phone cannot read them. Windows has the OpenAI fix since 1.9.0 (16 October); the Claude values from Claude Code and the same sessions card come there with 1.10.0.

In the Mac App Store since 16 October 2026 — a free update. If you skipped 7.5 or earlier, what they brought is further down this page.

Clear and calm: every card fits its text, outages show once, and nothing fails quietly any more.

Changed in 7.6

• Card header in two rows. Name and plan badge stay on top; the age of the figures and any notes move to a second row when space runs out. Nothing gets cut off any more — before, an outage note could shrink to “⚠ …”.
• One outage bar instead of three copies. When Anthropic has an incident, you used to see it on the Claude card, the second Claude card and the Anthropic API card. Now a single line under the summary names the provider, the kind of outage and every card it affects, with a link to the status page: “Anthropic: Partial outage — affects Claude, Claude 2, Anthropic API · Status page”. The cards keep only the warning sign, open or closed. The outage texts are now our own, in English and German — “Partial outage”, “Major outage”, “Critical outage” — instead of whatever the provider wrote. Copilot (GitHub) joins the bar.
• Calmer cards. The reset bell and the ⓘ button appear when you point at a window; a bell you have set stays visible. One forecast line instead of three: “Lasts until the reset · today’s share 14%”. The details sit behind ⓘ.
• A tidier header. Refresh and Settings stay in the panel header; everything else moved into a “…” menu, with labels instead of bare icons.
• Sparklines with a time axis. The line now spans the whole window, from its start to the reset, with a dashed plan line from 0 to 100%. You see at a glance whether you are ahead of plan or behind it.
• One number format. In English “63%”, in German “63 %” with a space that never breaks the line. Amounts read “USD 12.40” everywhere — on cards, in widgets, in the menu bar.
• Today’s share, said plainly: “Today’s share: 14% of the week, 0% used”.
• Settings. Two new pages, General and Costs, sort what used to be spread across other pages. Search jumps to the right section and highlights it. A limit or a credit you type in is saved at once, without a separate button. ⌘W closes the window; Return no longer does, so you no longer close it by mistake while typing.
• Status colors you can read. Green, orange and red meet a contrast of 4.5:1 in every theme, light and dark.
• Tiles carry your own account names, as the cards do. Plan badges show only plans the app knows — no more raw code names such as “Default raven”.
• Widgets. The ring tile shows the name of its source; the small list no longer runs over; values no longer shrink to fit; the ring follows tinted and clear modes; the Control Center control shows how old its figure is; today’s share follows the system language.

Fixed

• Folder access no longer vanishes. One failed attempt to open the saved access to your .claude folder used to delete it for good — then the sessions card and project usage were gone until you granted the folder again. Now the app keeps the access and asks you to renew it. The folder dialog starts in your home folder, and if you pick a folder one level too deep, the app corrects it.
• “No active session” was wrong when folder access was missing. The sessions card now says “Access missing”, and the usage window offers the button to grant it too. After a reinstall, a hint tells you how to get project usage back.
• History survives a reinstall. Usage history and the usage archive now live in the app group container, which macOS keeps when you delete and reinstall the app. Existing data moves over once, on the first launch. Settings › History can export all of it as one file and import it again — on the same Mac or a new one.
• OpenAI daily costs were grouped by your local day instead of by UTC, as OpenAI counts. The first day of a month could fall out, and the cost spike alarm could fire for no reason.
• The budget warning stayed silent all month when one cost card kept failing. Now it works with the cards that answer.
• A write error could count usage twice; a broken window_minutes value in a Codex file could crash the app; a time stamp from the future blocked a history series.
• VoiceOver reads the whole card header, and the steppers, fields and sliders in Settings.

New and moved settings

Good to know

• On first launch the history moves. 7.6 moves your usage history and archive into the app group container once. You need to do nothing; it happens at the first start after the update.
• After a reinstall, grant the .claude folder again in the sessions card. macOS does not keep folder access across a reinstall. Once you grant it, project usage comes back.
• Settings › History can export and import. One file holds your history — handy before a reinstall or for a new Mac.
• The settings have new pages, General and Costs. If you miss a switch, search finds it and jumps there.
• 7.6 adds no new address and sends nothing anywhere. The outage bar reads the same public status pages as before. Still free.
• Still runs from macOS 14.
• iPhone, iPad and Apple Watch have the same theme since 2.17. Windows stayed on 1.8.1 until 1.9.0 (16 October), which carries the OpenAI fixes.

In the Mac App Store since 13 October 2026 — a free update. If you skipped 7.4 or earlier, what they brought is further down this page.

Your money and your time in hand: a note before a plan renews, a bell that tells you when a window is free again, a daily share so the week lasts, quiet hours at night — and once a month, what AI cost you.

New in 7.5

• A renewal calendar. Under Settings › Display › Subscription costs there is a new part, Renewal calendar: for Claude, the second Claude account and ChatGPT/Codex you enter the next date a plan renews and whether it renews monthly or yearly. The app moves the date on by itself — a plan from the 31st renews on the last day of shorter months. Some days before (3 by default, 1 to 7) a notification comes at 9 AM: “Claude Max renews on 14 Oct (CHF 200).” For Claude it adds the plan check from 7.4 when it suggests another tier. From 7 days before, the card shows a line: “Renews in 5 days (14 Oct)”. Leave the date empty and nothing happens.
• A reset alarm. A bell sits in the title row of every window with a known reset — Claude 5 hours, week and model windows for both accounts, Codex and Kimi Coding. Click it and it fills: you get one notification when the window resets, or at once if the provider resets it early. A second click removes it. The alarm survives a restart and comes through in quiet hours and during a pause, because you asked for it.
• A daily share for the weekly window. Under every weekly window the card says how much you can still use today so the week lasts: “9 % left for today so the week lasts”, “4 % over today’s share” or “Today’s share reached”. It splits what is left at the start of the day across the days until the reset. Also in the large desktop widget, where there is room. Switch: Show daily share under Settings › Display › Quotas, on by default.
• Quiet hours. Settings › Notifications › Quiet hours: from and to (22:00 to 07:00 by default, across midnight), off until you switch it on. In that time the app stays silent. Let cost alerts through — on by default — lets the cost alert and the budget alert pass.
• Pause until tomorrow. A moon button in the panel header and an entry in the right-click menu of the menu bar icon mute every notification until quiet hours end, or until 7 AM. A band at the top of the panel says “Notifications paused until 07:00 · Resume”. Reset alarms still come.
• “Your AI month”. The app keeps a small month book on your Mac — subscription fees, API costs and weekly peaks, for up to 13 months. On the 1st, from 9 AM, a notification says what the month before cost: “Your AI month September: CHF 312 — plans 220, API 92.” The view shows the total, plans and API apart, each service, the month before, your average weekly use and — on the Mac only — the value of your Claude use at API list prices. Open it with the calendar button in the panel, from the right-click menu, by clicking the notification or through aicockpit://monat. Share as image puts “Claude” and “Claude 2” where your account names would be, with aicockpit.info at the bottom. In the first month it says honestly: “No full month recorded yet.”

New settings

Changed

• Notifications also appear while AI Cockpit is in front. Before, they stayed hidden while the app was active.
• Every notification passes one check for quiet hours and the pause — thresholds, forecast, sessions, off-schedule resets, cost alert and budget alike.

Fixed

• OpenAI costs are back. OpenAI now returns costs for the past year only and turns down any older query with HTTP 400. AI Cockpit asked for three years, so the card stayed empty. Now it reads one year (364 days); if OpenAI shortens the span further, the app asks once more from the date OpenAI names. When a service turns a query down, the card now shows its reason too, not just the number.
• The card of the second Claude account came back. After a reinstall the card was missing, although the account stayed signed in. A new switch, Show second Claude account, sits in Sections in the window, and the card returns by itself at launch when the account is signed in.

Good to know

• You enter the renewal date yourself. No provider tells an app when your plan renews, and AI Cockpit asks none of them.
• The daily share is a guide, not a limit. It only covers weekly windows and needs a known reset; on a full window it says nothing.
• Month book, alarms and quiet hours stay on your Mac. 7.5 adds no new address, sends nothing anywhere and collects no data. Still free.
• Still runs from macOS 14.
• iPhone, iPad and Apple Watch have the same since 2.16, Windows since 1.8.

In the Mac App Store since 10 October 2026 — a free update. If you skipped 7.3 or 7.2, what they brought is further down this page.

For developers, and for clarity: your limits in the Claude Code status line, a notification when a session is done, a tip on where to go when a window fills up — and a plain answer to “where do I stand?”.

New in 7.4

• Your limits in the Claude Code status line. Once you switch it on, AI Cockpit writes a small file after every refresh, ~/.claude/aicockpit/stand.json: percentages, reset times, amounts and a traffic-light level. A bundled script, statusline.sh, sits next to it and turns it into a line like “Claude 63 % · Codex 49 % · OpenAI $2.87”. Copy command hands you the statusLine block for Claude Code’s settings.json. If the script moves — because you granted or revoked the folder — the settings say “Path changed – copy command again”.
• “Claude is done.” A notification when a Claude Code session has finished its work: after a run of at least a minute, 30 seconds of quiet, and a reply that does not end with a question. It sits next to “waiting for you”, never both at once, with a switch of its own. After the update it follows your “waiting for you” setting.
• A tip on where to go next. When a window reaches 90 % and Codex, Claude or a second Claude account still has more than half left, the card says so: “Codex has 70 % left (5 h)”. Never to an API that charges per call.
• A plan check from your weekly use. AI Cockpit keeps the weekly peak of the last four weeks and sets average and peak against your Claude plan — “Max 5× would likely do”. A verdict comes after two full weeks, each watched by the app for at least three days. Codex shows average and peak only. It sits next to the API-value check from 7.3.
• One sentence on where you stand, at the top of the panel and in the menu bar tooltip: “All in the green”, “Claude week runs out Thu 4:40 PM”, “Monthly budget almost reached”. It uses your warning threshold, so it matches the card colours.
• (i) buttons that explain the terms — next to the 5-hour, weekly and model windows, the credits, the time marker, “above plan” and the alert settings. A click opens two or three sentences.

New settings

Good to know

• The plan check is an estimate. Providers publish no fixed weekly limits, and the app only sees the peaks it fetched. The verdict covers Claude only.
• Switching the status off removes file and script from every place the app may have put them. The status line then shows nothing; switch it on again and copy the command anew.
• The status file stays on your Mac. 7.4 adds no new address and sends nothing anywhere.
• Still runs from macOS 14; the control for Control Center needs macOS 26, as in 7.3.
• The tip, the sentence, the plan check and the (i) buttons come to iPhone, iPad and Apple Watch with 2.15. Windows has all of 7.4 since 1.7 — the status line and “Claude is done” included.

In the Mac App Store since 8 October 2026 — a free update. If you skipped 7.2, what it brought is further down this page.

No surprises: a notification when a window resets off schedule or your API spend jumps, a check on whether your Claude plan pays off, a control for Control Center and your status as an image — plus three colour schemes and a colour of your own for each card.

New in 7.3

• An alert when a window resets off schedule. Providers sometimes reset a window before the announced time — Codex, for one, after an outage. That is quota you get back without ever noticing. Now a notification says so: “Codex reset ahead of schedule — full quota, you got about 40 % back.” To rule out a glitch, it waits until the next fetch confirms the low value or the provider names a new reset time. On by default.
• A cost alert. A notification when credit at OpenRouter, DeepSeek or Kimi drops fast within an hour — by at least the amount you set and by a multiple of your hourly average over the last 7 days. For the OpenAI and Anthropic APIs it comes when today’s cost runs at least twice as high as on a typical day. At most every 6 hours per credit, once a day per API.
• “Is your subscription worth it?” The Usage window sets the value of your Claude Code use at API list prices against the price of your Claude plan, and rates it: clearly worth it, worth it, borderline or probably not. If a cheaper tier would have done, it says so — “Pro would have been enough”. Pick Pro, Max 5× or Max 20×, or enter your own amount, for each Claude account. With less than 7 days of data it warns that the projection is still rough.
• A control for Control Center (macOS 26 or later). It shows the tightest value and fetches everything again on click.
• Share your status as an image. A menu in the panel shares it, copies it to the clipboard or saves it. Claude account names are hidden, and a small aicockpit.info line sits at the bottom.
• Three colour schemes. Graphite is a neutral grey, Deep Blue a cool navy, Classic the colours up to 7.2 — each dark or light, or as the system. The scheme covers the window, the menu bar and the widget. Every text colour was checked for a contrast of at least 4.5:1, against the card and against the background.
• A colour for each card. One of twelve tested shades or any colour you like, with a way back to the default — handy to tell two Claude accounts apart.

New settings

Good to know

• OpenAI and Anthropic report costs late, often by hours, so their alert can come well after the spend. Credit at OpenRouter, DeepSeek and Kimi is read with every refresh — that alert comes within one interval.
• The subscription check is a guide, not a bill. It counts Claude Code use at API list prices (as of 8 October 2026), for Claude only, and needs the plan price you enter. Nothing is sent anywhere for it.
• Off-schedule resets and cost alerts are worked out on your Mac, from figures the app already fetches. 7.3 adds no new address.
• The control needs macOS 26. Everything else in 7.3 still runs from macOS 14.
• Both alerts, the control, the image and the colour schemes are on iPhone, iPad and Apple Watch too, since 2.14. The subscription check stays on the Mac. Windows stays on 1.6 for now.

In the Mac App Store since 7 October 2026 — a free update. 7.2 takes the place of 7.1, which was withdrawn before release, and includes everything from it, from 7.0 and from 6.9 — none of them came out on its own. Below: what 7.2 brought, coming from 6.8.

Figures that are honest about their age, a forecast that says when a window runs out, and fewer false “signed out” messages — plus everything waiting since 6.8: a new Accounts page, DeepSeek as the ninth service, settings you can search, numbers you can read either way, a time marker and a widget you set up.

New in 7.2

• A forecast that needs no history. Under every quota window: “At this pace full at 16:40 — before the reset”, or “lasts until the reset (≈ 70 %)”, worked out from the time marker alone. It steps in right after launch or with history switched off, until the forecast from your history has enough data — and it warns by notification too.
• Wakes up with your Mac. After sleep the app refreshes as soon as the network is back, instead of waiting for the next interval.
• A right-click menu on the menu bar icon: Refresh, Settings, Usage, Quit.
• The second Claude account everywhere: in the menu bar, in the forecast notification and in the diagnostics. The key figures bar now also offers Kimi, OpenRouter, Grok, Copilot, DeepSeek and the second Claude account.
• GitHub outages on the Copilot card, read from www.githubstatus.com — as for Claude, OpenAI and Kimi.
• Clicking the widget opens the panel.
• Notification permission in the settings: you see whether macOS lets the app notify you, and the way to System Settings when it does not.
• Usage window: group projects by their top folder, or show the full path.
• xAI team ID by hand, for management keys without team access. And the Claude card can be hidden if you do not use Claude.

Changed in 7.2

• Throttling no longer means “signed out”. When Claude or Codex renews a sign-in, the app now tells throttling (429), server errors, timeouts and a real rejection apart, and renews at most once per round. HTTP 403 is either throttling (GitHub) or a missing permission — the figures stay on the card instead of being wiped.
• A rejected key rests for an hour instead of being tried again every interval.
• Anthropic costs load faster. Completed days (UTC) are kept; each refresh fetches only the rest instead of three years in about 36 pages.
• Old figures look old. Collapsed cards and tiles show “⚠ 2 h ago” when a source keeps failing; the header no longer counts the sessions clock. A reset in the past reads “reset — updating” and fetches anew.
• The widget warns where the app warns. It uses the app’s warning levels and thresholds, shows the age per row, puts the ring on the most pressing window, reloads once its data is 40 minutes older, and drops what no source supplies any more.
• Menu bar: grey when its source is failing, and a short tag when it falls back to another service.
• Intervals: at least 60 seconds for Claude and Codex; a change applies after one second.
• Key fields trim spaces and warn about a wrong key prefix; a failed write to the keychain is shown.
• The settings search finds every control, the menu bar styles included. “Left” now also applies to the day line, the tiles and the OpenRouter summary.
• Cost report and daily curve count by UTC, like the providers; on the 1st the monthly sum no longer carries last month’s values.

Fixed in 7.2

• The “reset” notification never fired. It does now, and notification levels survive a restart.
• Windows and links opened behind the floating panel.
• A made-up “Opus” weekly line could appear next to “Fable” from an old fallback key.
• Codex sign-in: the app server recovers after a failed start, a cancelled sign-in frees its port at once, and answers without a state are refused.
• Possible crashes with duplicate session IDs and odd Copilot numbers.

From 7.1

7.1 was withdrawn before release; all of this arrives with 7.2.

• A new Accounts page. One row per service: its coloured letter as on the cards, the name in large type, one line on what it brings (plan tier, API costs, credit balance) and a status pill — “active”, the last four characters of a stored key, “expired” or “error” in orange, “missing” in grey. Subscriptions with a sign-in and API keys form two groups; the services you have set up come first.
• A page per service. Each row opens its own page: sign in or enter the key, replace or remove it, what it does, and where to get it — with a link to the provider’s page. The usual pitfalls — where an admin key is needed, the management key for Grok, Copilot for personal plans only, the right Kimi platform — stand out as a highlighted note instead of fine print.
• Search jumps straight to a service. A settings search that matches exactly one service — “deepseek”, say — opens its page directly.
• A short weekly reset in the widget. “↻ Thu 4:13 PM”, and from a week out just the date, instead of “Reset” followed by a date that got cut off.

From 7.0

7.0 never reached the Mac App Store on its own; all of this arrives with 7.2.

• DeepSeek, the ninth service. Your DeepSeek balance, with what you topped up and what was granted. Add an ordinary API key from platform.deepseek.com under Settings → Accounts; the card fetches every 5 minutes and only reads.
• Two currencies stay two. If your account holds dollars and yuan, each shows as its own line — never added up.
• An empty balance is named. When DeepSeek stops requests because the balance has run out, the card says so instead of showing an error.
• Search the settings. A search field at the top of the settings sidebar finds every switch, in English and German at once — whichever language the app runs in — and ignores case and accents: “ubersicht” finds “Übersicht”. Several words must all match, but not in the same term. Enter jumps to the first match, Esc clears the field, ⌘F puts you back in it.

New since 6.8 — from 6.9

6.9 never reached the Mac App Store on its own either; all of this arrives with 7.2.

• Used or left. A new choice under Settings → Display → Quotas shows every quota as used (the default, as before) or as what is left — on cards, bars, collapsed summaries, key figures, the menu bar (every icon style; the ring fills with the number shown) and the widget. Where there is room, the word “left” stands next to the number, also in the menu bar tooltip.
• Warning colours still follow usage. 90 % left stays calm, 5 % left turns red. Only the number flips, never the traffic light.
• Both views always add up to 100. 71.5 % used reads as 72 % used — or 28 % left, not 29 %.
• The widget can be edited. Right-click, Edit “AI Cockpit”: pick one source or all cards, and whether it counts as in the app, used or left.
• Codex credits in the large widget, on a line of their own — as credits, never converted to dollars.
• A time marker in the bar. A line with a small triangle shows how much of the window has passed. If 3 of 5 hours are gone, it sits at 60 %: a bar at 45 % means under plan, at 75 % over. The reset line says it in words — “Resets in 1 h 42 min · 18:42 · 15 % over plan”, “on plan” or “under plan” — and the menu bar tooltip adds the same. The marker moves every minute, in the widget too. With “left” it moves to the other side. Switch it off under Display → Quotas (“Time marker in the bar”, on by default). Money cards and the context window have no marker.
• A live countdown in the widget. When the reset is less than 24 hours away, the large widget and the ring of the small one count down every second (“↻ 1:42:07”); for weekly windows the clock time stays.

Widget options

Also changed since 6.8

• Calmer refresh intervals. Claude and ChatGPT/Codex every 3 minutes (before: 1 minute and 30 seconds), all other cards every 5 minutes (before: up to 15). Only intervals still on the old default move; your own choices stay. Refresh now still fetches at once.
• Slower after repeated throttling. If a service answers 429 twice within an hour, that card is checked half as often (3 → 6 → 12 minutes, at most 15) and speeds up again one step per quiet hour. Retry-After is still honoured.
• Cards no longer all ask at the same second, for example after waking from sleep: each fetch waits a random 0–20 seconds. Sessions and the cost check are left out — they read local files.
• The large widget keeps its height. With many sources plus the credit line it leaves out a card and shows “+ N more” instead of spilling over the edge.

Also fixed since 6.8

• The widget showed GitHub Copilot with the letter of another service. It now uses “P”, as the app does.
• The Usage window follows a switch between light and dark while it is open. Before, it could keep grey text on a light background.

Good to know

• DeepSeek offers no endpoint for usage, only the balance. The card shows what is there and says what is missing — no invented spend.
• The day line, warnings, the report export, sparklines and the forecast keep showing usage — they describe consumption.
• The context window on the sessions card always reads as used: it is not a quota.
• The forecast from your history stays; the new one fills in until there is enough of it. Both are a projection from your pace, not a promise.
• www.githubstatus.com is the one new address. Nothing is sent there; only GitHub’s public status file comes back.
• The same Accounts overview is on iPhone and iPad since 2.12 (in the App Store since 3 October 2026) and on Windows since 1.6. The forecast line is on iPhone too, since 2.13 (in the App Store since 3 October 2026).
• Still runs from macOS 14.