Reset

Reset is a free, open source macOS app for people who live inside Claude Code (with Codex, Gemini CLI, Antigravity and Cursor support too). It sits in your menu bar, counts down to your next fresh window, and turns your local usage logs into a dashboard, desktop widgets and a card you can post.

Everything it reads stays on your Mac.

What it does

• Your window at a glance. A colored pill in the menu bar: orange "2:46 · 42%" as usual, red when the window is nearly empty or your pace runs out before the reset ("38% · out 10:07": 38% left, empty at 10:07 at this pace), green "ready" when a fresh window is waiting. Exact when live usage is on; estimated from your activity when it isn't, and labelled as such.
• Phone alerts while your Mac sleeps. Reset queues each reset alert on an ntfy server ahead of time, so it arrives on your phone even with the lid closed.
• Limit forecast. At your current pace, will you hit the limit before the reset? Reset warns you once per window, on your Mac and your phone.
• Where your window went. A dashboard card (and a share card) showing how much of your usage re-did work it didn't need to: re-reading long sessions after their cache expired, context rebuilt mid-session, the same tool failing again and again. Each with a tip; on one Mac it was a fifth of a month, almost all from coming back to big sessions after a break. Claude can ask for it too (get_waste_report).
• Break Guard. When a big session goes quiet, the menu says how long its cache stays warm and what /compact costs now against re-reading it later; the menu bar counts down the last 15 minutes, and a heads-up reaches your phone before it expires. After a long break, start fresh, don't resume.
• Weekly limit, early. On Pro and Max plans, Reset warns from 80% of the weekly limit (in the menu, the status line and one notification a week), so a used-up week doesn't come as a surprise. To keep an eye on it all the time, Settings → Appearance → Limits shows the weekly limit under the 5-hour window in the menu bar pill ("Both"), or on its own ("Weekly"). For Claude this needs live usage.
• Work and personal accounts. Use a Team account and a personal Pro plan? Each gets its own entry, window and countdown, and the dashboard splits costs, sessions and projects by account, each against its own plan price. The menu's top card, the menu bar and the mascot follow the account you're using. Live usage covers the account Claude Code is signed in to and, if you allow it, the other accounts signed in to the Claude desktop app.
• Notifications you can act on. Clicking one opens the dashboard; reset alerts have a Start now button and sign-in problems a Sign in button. Times follow your 12 or 24-hour clock.
• Auto-start. Optional. When a window runs out, Reset sends a one-word prompt so the next window starts right away instead of when you wake up.
• Dashboard. API-price value of your usage against your plan, cache savings, every 5-hour window of the week, token mix, when you build, cost per session and per project, and what used each session (main thread, subagents, each MCP server). Right-click a session to copy the command that resumes it; with two accounts it says which one the session ran under, and warns when Claude Code is signed in to the other. ⌘1–4 switch the range, ⌘R rescans.
• 8 desktop widgets, including Fuse: your window as a burning fuse, with where today's pace runs out and how much to slow down to make the reset. The others: countdown, gauge, arrival, today's spend, month calendar, week, and a Game Boy. When a fresh window is waiting, the countdown and Fuse widgets show a Start now button that starts it right from your desktop. Plus 8 looks for the menu card, and the other tools and accounts can each get a card too (Settings → Appearance).
• Mascots. Animated desktop pets whose mood and health bar follow your usage: fresh, focused, tired, asleep at the limit, waking at the reset.
• Share card. A 1200 × 675 image of your month for X.
• Claude Code integration. A status line for your terminal and an MCP server, so Claude can check how much window is left (and the weekly limit) and pace itself, including a per-task budget that tells it when to wrap up or switch to a smaller model. One click in Settings sets both up, and both say so when Reset isn't running rather than showing old numbers.
• Stay awake. Keeps the Mac from idle-sleeping (when plugged in, or always) so auto-start and alerts run on time, without a separate app. The display still sleeps.
• One-minute setup. A first-run guide picks your tools, signs Claude Code in, sets up phone alerts with a QR code, and takes your plan price.

Install

Download: grab the latest .dmg from Releases, open it and drag Reset to Applications.

Reset isn't notarized yet, so macOS blocks the first launch. Open Reset once, close the warning, then go to System Settings → Privacy & Security and click Open Anyway next to Reset. You only do this once. (Or run xattr -dr com.apple.quarantine /Applications/Reset.app in Terminal.)

The download is ad-hoc signed, so its widgets won't appear. For widgets, build it yourself as below.

Build from source (about two minutes):

You need macOS 14 or later, Xcode 16 or later, and XcodeGen:

brew install xcodegen
git clone https://github.com/msfbeast/reset.git
cd reset
./scripts/build.sh --install

That builds an ad-hoc signed app and copies it to /Applications.

Widgets need a real signature. macOS only lists widgets from apps signed with an Apple Development certificate. If you're signed in to Xcode with any Apple ID (a free one works), pass your team ID:

TEAM=ABCDE12345 ./scripts/build.sh --install

Find your team ID in Xcode under Settings → Accounts, or in Keychain Access on your "Apple Development" certificate.

Setting it up

The first launch opens a short setup guide that covers all of this. You can run it again from Settings → Advanced, or change anything below in Settings, which has four tabs: General (live usage, your tools, alerts), Automation (auto-start, stay awake), Appearance (look, mascots) and Advanced (commands, custom tools, Claude Code integration, diagnostics).

Feature	How
Exact Claude usage	Settings → General: turn on Live usage from your Claude account. Reset reads the sign-in that the Claude Code CLI stores in your Keychain and asks Anthropic for your current usage every 2 minutes while you're using Claude Code, and every 5 minutes otherwise. If it says the sign-in expired, run claude auth login once.
Phone alerts	Install the free ntfy app, subscribe to a long random topic name, and enter the same topic under Settings → General → Phone alerts. Anyone who knows the topic can read it, so make it unguessable.
Auto-start	Settings → Automation: turn on Auto-start for Claude. It runs claude -p "hi" --model haiku (the command is under Advanced) once when a window you were watching runs out, while the Mac is awake. Each run uses a little of your quota.
Value multiple	Press Add your plan price on the dashboard and enter what you pay per month.
Stay awake	Settings → Automation → Stay awake: Off, When plugged in, or Always. Blocks idle sleep only; closing the lid still sleeps the Mac unless it's plugged in with an external display.
Status line	Settings → Advanced → Claude Code integration → Add for me (it backs up settings.json first), or add to ~/.claude/settings.json: "statusLine": {"type": "command", "command": "cat \"$HOME/Library/Application Support/Reset/statusline.txt\""}
MCP server	Add for me in the same place, or claude mcp add --scope user reset -- /Applications/Reset.app/Contents/MacOS/reset-mcp (with Reset's real location if it isn't in Applications). Tools: get_claude_usage, plus set_budget (task, max window %, optional downshift model) and clear_budget. With a budget set, the advice moves from normal to wrap-up at 80% of the cap to downshift at the cap. The active budget shows under the burn bar on the dashboard, with a Clear button. It's kept in budget.json next to widget.json and never leaves the Mac.

FAQ

Does Claude checking its own window over MCP use up that window? Barely. Reset's MCP server makes no API call; it reads a file Reset already keeps on your Mac. The only cost is the answer landing in the conversation, about 150 to 250 tokens a check, plus about 350 tokens for the three tools' descriptions in context. That's less than Claude reading one small file.

Can an agent tell when the numbers are old? Yes. get_claude_usage includes usage_as_of and usage_age_minutes (when the usage was last read from Anthropic), and if the live check starts failing, live_check_failing_since_local, the reason, and usage_may_be_stale: true. The advice and any budget then say the reading may be out of date. If Reset itself quits, everything is marked stale.

How exact is the countdown? With live usage on, the reset time and % come straight from your Claude account, refreshed every 2 minutes while you work and every 5 when idle. Without it, Reset estimates from your Claude Code activity and labels the numbers "estimated", so you always know which you're looking at.

Does it track the weekly limit? Yes, on Pro and Max plans with live usage on. The menu shows the weekly %, it warns you from 80% (in the menu, the status line and one notification a week), and once it's used up it shows when it lifts. Claude sees it through the MCP server too.

How do task budgets work? Before a big task, Claude (or you, by asking it) sets a budget over MCP: a task name, the window % to stop at, and optionally a smaller model to fall back to. As the window nears that % Claude is told to wrap up, and past it to move to the fallback model or stop and check in. Reset doesn't switch models itself, and Claude can't change its own: it hands the remaining steps to subagents on that model (which start fresh, with only the brief Claude writes), or asks you to run /model to continue with the whole conversation. The cap is the window's total usage ("stop when the window hits 60%"), not the task's own share. It's guidance Claude follows, not a hard lock.

Does auto-start use up my usage? Almost none. It sends one word to the smallest model (Haiku) with no tools or MCP servers loaded, which is just enough to start the 5-hour clock. It waits for a working network instead of failing offline, and gives up after 4 tries.

Does it work with the lid closed? Phone alerts do: each one is queued on the ntfy server ahead of time, so it arrives while the Mac sleeps. Auto-start needs the Mac awake, so leave it plugged in with Stay awake on; on battery with the lid closed, it starts the window the moment the Mac wakes.

I have two Claude accounts. Each gets its own entry, window, costs and plan price, and you can name them (Settings → General → Claude accounts). Live usage, the forecast, auto-start and budgets follow the account Claude Code is signed in to; the other one is estimated. See Troubleshooting for details.

How does it split costs between two accounts? Claude Code records which account each session ran under, but only since early October 2026. For older sessions Reset works it out: Claude desktop app sessions go to the account named in their folder, and Code-tab sessions from before a second account appeared in the desktop app go to the one that was there. Otherwise the desktop app's own usage history decides, and anything left goes to the account Claude Code is signed in to. It's a well-supported guess for those older sessions, not a record, so their split can be slightly off; everything since early October is exact.

How is "where your window went" worked out? From token counts and times only. A break is a turn that had to write a session's whole context to cache again because its cache had expired (an hour after the last turn with Claude Code's usual cache); it's an upper bound, since it assumes you'd have carried on in that session. A bust is a turn that reused under half of what the previous one had cached (compactions don't count). A loop is the same tool failing three times or more in a row. "Opus where Sonnet would do" is an estimate and isn't added to the total. Dollars are API list prices; on a plan it shows a share of your usage instead.

Does it read my prompts or code? No. From Claude Code's local logs it keeps token counts, model names, times, tool names, project folders and session titles, never the text of your prompts or Claude's replies. See Privacy for exactly what leaves your Mac and when.

Is this made by Anthropic? No. Reset is an independent, open source project, not affiliated with or endorsed by Anthropic.

Why does the first launch need "Open Anyway", and why no widgets in the download? Reset isn't notarized by Apple yet. Notarization is coming; until then, open it once via System Settings → Privacy & Security, and build from source for widgets (see Install).

Troubleshooting

macOS says it can't open Reset. It isn't notarized yet. Open it once, close the warning, then go to System Settings → Privacy & Security and click Open Anyway. You only do this once.

I don't see Reset's widgets. The download is ad-hoc signed, and macOS only lists widgets from apps signed with an Apple Development certificate. Build it yourself with your team ID (see Install); a free Apple ID works. If they still don't appear, log out and back in so macOS refreshes the widget list.

The menu or widgets say "estimated". Reset is guessing from your Claude Code activity instead of reading your real usage. Turn on Settings → General → Live usage from your Claude account. The note says why it's guessing:

• live usage off: turn it on.
• Claude Code signed out: press Fix in the menu, or run claude auth login in Terminal.
• waiting for live usage: it was just switched on or Reset just started; give it a minute.
• live usage unavailable: Settings → General shows the reason under the switch.

If macOS asks whether security may read "Claude Code-credentials" from your Keychain, click Always Allow. That's how Reset reads Claude Code's sign-in.

The countdown is off, or the window started earlier than Reset thinks. Without live usage, Reset can't see Claude used on claude.ai or in the desktop app, only Claude Code. Live usage fixes this. Without it, pick I just used Claude on the web from the menu's … button when you start there.

I use two Claude accounts (say, a work Team plan and a personal Pro plan). Reset shows each as its own entry, with its own window and countdown, worked out from that account's Claude Code sessions. The menu's top card, the menu bar and the mascot follow the account you're using: switching accounts in the Claude desktop app moves them. Live usage, the forecast, auto-start and the widgets go through the account Claude Code in Terminal is signed in to; with live usage on, Reset asks Anthropic which account that is, because claude auth status can show the Claude desktop app's account instead. Other accounts are estimated from their sessions and the desktop app's readings, unless they're signed in to the Claude desktop app: then turn on Settings → Advanced → Live usage for other accounts in the Claude desktop app for live usage there too. The MCP server answers for the account the asking agent runs on. The dashboard's All / Team / Pro switch splits costs, sessions and projects by account, and each account can have its own plan price (Settings → General).

The menu bar shows only the countdown, with no percentage. There's no usage reading yet, and Reset won't show the clock as if it were usage. Turn on live usage. Want plain text instead of the colored pill? Settings → Appearance → Menu bar.

Phone alerts don't arrive. Check that the topic in Settings → General → Phone alerts matches the one you subscribed to in the ntfy app exactly, then press Send test. Make sure notifications are allowed for ntfy on your phone, and to get through Do Not Disturb, allow high-priority alerts for that topic in the ntfy app.

Mac notifications don't show. Allow them in System Settings → Notifications → Reset.

Auto-start didn't run. Settings → Automation shows each tool's last run and why it failed, with a Sign in button if Claude Code was signed out. Also know that auto-start:

• only runs for a window Reset watched run out, never right when you switch it on;
• is skipped while a weekly limit is used up;
• needs the Claude Code command-line tool installed and signed in;
• runs late if the Mac was asleep. Settings → Automation → Stay awake keeps it awake when plugged in or always.

Settings shows a "Claude Code log format" warning. A Claude Code update probably changed how it writes its logs, so some usage can't be read. Check for a Reset update in the same box, or report it.

Something else. Settings → Advanced → Report a problem opens a GitHub issue with versions and on/off states filled in, never usage numbers, paths, project names or your phone topic. You see everything before sending.

Uninstalling. Quit Reset, drag it to the Trash, and delete ~/Library/Application Support/Reset. If you set them up, also run claude mcp remove --scope user reset and remove the statusLine entry from ~/.claude/settings.json.

Privacy

Reset reads local files only:

• Claude Code transcripts in ~/.claude/projects. Reset scans them for usage and keeps only token counts, models, timestamps, tool and MCP names, project folders and session titles, in a cache in ~/Library/Application Support/Reset. Your prompts and Claude's replies are never stored.
• ~/Library/Application Support/Claude/plan-usage-history.json from the Claude desktop app
• Only if you turn on Live usage for other accounts in the Claude desktop app (Settings → Advanced): the sign-in the desktop app keeps in ~/Library/Application Support/Claude/config.json, decrypted in memory with the desktop app's key from your Keychain. macOS asks you first; click Always Allow so it doesn't ask again each time Reset starts. The downloaded build is ad-hoc signed, so macOS asks once more after each update. The same key can decrypt other data the desktop app encrypts on this Mac; Reset reads only the sign-in
• Codex session logs in ~/.codex/sessions
• Antigravity's local language server, while Antigravity is open
• If you switch Cursor on: the Cursor app's sign-in and plan name in ~/Library/Application Support/Cursor/User/globalStorage/state.vscdb, opened read-only

It only sends anything off your Mac for these features, and only to the place you'd expect:

• Live Claude usage sends Claude Code's sign-in to api.anthropic.com to read your usage and which account it belongs to. With Live usage for other accounts in the Claude desktop app (Settings → Advanced) on, it does the same with the desktop app's sign-in for each of its other accounts. Reset never stores, refreshes or rewrites a sign-in.
• Cursor, off until you switch it on in Settings, sends the Cursor app's sign-in to api2.cursor.sh to read this month's usage. Reset never stores, refreshes or rewrites that sign-in.
• Phone alerts send the alert text to the ntfy server you choose. The text is only reset and run-out times, such as "Claude has reset: fresh window ready since 7:40 PM", but anyone who knows your topic name can read it, so pick a long, random topic.
• Update checks ask api.github.com once a day whether a newer release exists. Turn this off in Settings → General.

No analytics, no accounts, no servers of its own.

About the MCP server: when Claude calls one of Reset's MCP tools, the result goes into your Claude Code conversation like any other tool result, so Anthropic sees it: your window usage %, reset time, API-value cost, and the task name of any budget you set. Don't put anything in a budget's task name you wouldn't type in a prompt. The terminal status line is different: it's only shown in your terminal and isn't sent to the model.

How the numbers work

• Cost is the API list price of your tokens (input, cache writes split into 5-minute and 1-hour, cache reads, output, and advisor calls), not what you're billed on a subscription. Prices live in Reset/Usage/Pricing.swift.
• De-duplication. Claude Code writes one transcript row per streamed content block, and resumed sessions copy rows between files. Reset counts each response once, by message id, keeping the row with the final output count.
• Window estimate. Without live usage, a window is assumed to open at your first message after the previous one ended, rounded down to 10 minutes, and run 5 hours. Exact reset times from Claude Code's own limit messages override the estimate.

Tests

./scripts/test.sh

GitHub runs them on every push and pull request (all but the UI snapshots). The unit tests cover pricing, de-duplication, incremental parsing, window estimates, the forecast, project grouping, update checks, the MCP server's JSON-RPC round-trips and budgets, and the format-drift check below.

UI snapshots

./scripts/snapshots.sh            # compare
./scripts/snapshots.sh --record   # rewrite the references after an intended change

Draws the dashboard hero at three widths and in its main states, the menu bar pill and every widget look, with made-up numbers, and compares them pixel by pixel with the images in Fixtures/snapshots. A layout slip like the arc running through the countdown fails the suite; a failure saves the actual image and a diff (changed pixels in red) to the temporary folder. Rendering shifts slightly between macOS versions, so record the references on the Mac that runs the tests.

Format-drift check

./scripts/drift-check.sh                    # compare
./scripts/drift-check.sh --update-goldens   # rewrite the goldens from the current parsers

Fixtures/transcripts/ holds small anonymised samples of every log format Reset parses: a streamed multi-row Claude Code session, one that hits the 5-hour and weekly limits, a resumed session that copies rows into a second file, a Codex session with rate_limits, and the Claude desktop app's plan-usage-history.json. The check parses them with the real parsers and compares the result with Fixtures/golden/. It runs as part of ./scripts/test.sh too (DriftTests), and exits non-zero with a diff on any mismatch.

When the drift check fails

The fixtures are frozen files, so a failure means Reset's parsing changed, not that a tool changed its format.

• You changed a parser on purpose: read the diff the script prints (it compares Fixtures/golden with the fresh output in your temporary folder). If every difference is one you intended, run --update-goldens and commit the new goldens with the parser change.
• You didn't mean to change anything: treat it as a regression and fix the code, not the golden.

After a Claude Code update, Reset watches for you: while scanning, it counts recent transcript rows that look like usage but can't be read (bad JSON, a timestamp it can't parse, a missing message, id, model or token counts, token counts that aren't numbers, an unknown limit kind), and notices when Claude Code writes plenty of log lines with no usage in them at all. If that happens in the last 2 days, Settings shows a warning at the top naming the Claude Code version. Healthy logs show one quiet line under Settings → Advanced → Diagnostics. It runs on your Mac and sends nothing. The goldens themselves won't fail, because they don't read your live logs. To capture a new format for them:

1. Copy a few fresh rows of the changed kind from ~/.claude/projects/…/*.jsonl (or the other source).
2. Anonymise them: replace ids, uuids, session ids, request ids, cwd, org ids, and all prompt, reply, thinking and signature text. Keep every key and the shape and types of every value; that's what the parsers depend on.
3. Add them to the matching fixture, or a new one with its own test in ResetTests/DriftTests.swift, then run the check. A diff you didn't expect means the new format breaks Reset: fix the parser, then --update-goldens.
4. Before committing, grep Fixtures/ for your username, home folder and real ids.

Demo mode

For screenshots and screen recordings, start Reset with made-up data:

open -a Reset --args -demo YES

Nothing is read, written, alerted or auto-started while demo mode is on.

To see the log-format warning without breaking a real transcript, quit Reset and open Settings with a made-up problem:

open -a Reset --args -openSettings YES -simulateLogFormat problems

Use no_rows instead of problems for the "no usage found" variant. -renderSettings /path/to/settings.png saves a picture of the window instead.

Trademarks

Reset is an independent project and is not affiliated with, endorsed by or sponsored by Anthropic, OpenAI, Google or Anysphere. Claude, Codex, Gemini, Antigravity and Cursor are trademarks of their owners. The mascots are unofficial fan art based on those products' logos; you can turn them off in Settings → Appearance. If you represent one of these companies and want the fan art changed or removed, please open an issue.

Contributing

Pull requests are welcome. See CONTRIBUTING.md for what fits, the privacy rules and how to run the tests.

License

MIT. See LICENSE. It grants no rights to the logos the mascot fan art is based on; see NOTICE and Trademarks above.