Downloads

Every Glint release.

macOS 14 Sonoma or later, Apple Silicon. Grab the latest build below, or read the notes for any version.

v0.2.1Beta2026-07-06
Glint 0.2.1 Beta
3.7 MB download · 6.8 MB installed · macOS 14.0+ · ad-hoc signed
Install with HomebrewRecommended
$ brew install --cask binary-beam/tap/glint⧉ copy
Installs the latest build and opens without the Gatekeeper prompt. Update anytime with:
$ brew upgrade --cask glint⧉ copy
Or download the .dmg Download .dmg
First launch is blocked because this beta isn't notarized yet - go to System Settings → Privacy & Security, scroll down, and click Open Anyway. Still blocked? Run:
$ xattr -dr com.apple.quarantine /Applications/Glint.app⧉ copy
sha256 0d5c02801f71926538147dacd6b80f9148916d28d393ef039f8dfee277e30e19
Added
  • Automatic update checks (opt-in, Settings → Other → Updates). With the toggle on, Glint quietly looks for a newer version when you open the menu-bar panel - at most once every 6 hours, one tiny request, cached result, and still nothing running in the background (no timers, no polling). When an update is found, an "Update available" notice replaces the version line at the top of the panel; clicking it opens the downloads page. The manual "Check for Updates" button feeds the same notice.

All versions

  1. v0.2.1BetaLatest2026-07-063.7 MB.dmg
    Added
    • Automatic update checks (opt-in, Settings → Other → Updates). With the toggle on, Glint quietly looks for a newer version when you open the menu-bar panel - at most once every 6 hours, one tiny request, cached result, and still nothing running in the background (no timers, no polling). When an update is found, an "Update available" notice replaces the version line at the top of the panel; clicking it opens the downloads page. The manual "Check for Updates" button feeds the same notice.
    sha256 0d5c02801f719265
  2. v0.2.0Beta2026-07-053.6 MB
    Changed
    • Codex transcript parsing reuses cached date formatters (faster first read of a large session), and the menu-bar panel no longer runs the Claude CLI while the app is locked.
    Fixed
    • Reopening the editor (or Glint) while a session waited on a question no longer randomly flips it to "Thinking" or drops it entirely. Session files replay in arbitrary order on a relaunch, and a subagent's pre-question history could be mistaken for "the user answered" - subagent activity now only clears "Awaiting" when its lines are newer than the question itself.
    • Running a slash command (like /model or /help) while a session waits on a question no longer clears "Awaiting" and restarts the clock - only a real prompt or an actual answer un-blocks the session.
    • Cost estimates now cover Fable and Mythos models (previously $0.000), bill Haiku 4.5 and Opus 4.5+ at their correct rates, and keep the old rates for legacy Opus and Haiku models.
    • A usage limit whose reset text can't be parsed no longer keeps its session pinned "Limited" forever - it now falls back to the scope's maximum window (5h for session limits, 7 days for weekly).
    • An abandoned "Awaiting" session (terminal closed mid-question) no longer occupies the notch for a whole day - it stays visible for 10 minutes, then drops.
    • Kiro plan rows now complete when kiro-cli emits numeric task ids, and the Kiro account line parses regardless of the CLI's casing.
    • A hung "kiro-cli --version" can no longer wedge Kiro usage refreshes for the rest of the app run - every CLI spawn is now watchdogged, and CLI versions are cached so a usage refresh spawns fewer processes.
    • Elapsed times over an hour now read "2h 5m" instead of "125m 30s".
    • Revoking Accessibility permission mid-run now hides the Dock-side bar instead of leaving it frozen where the Dock used to be.
    • License activation now times out after 15s instead of hanging on "Activating..." for a minute when offline.
    sha256 0a63a832eed7b9d2
  3. v0.1.9Beta2026-07-023.6 MB
    Fixed
    • A session no longer falls to a false "Idle" while its subagents are still running. When the parent turn is orchestrating delegated agents its own transcript goes quiet and yields no turn-start line (and if Glint began watching mid-run, or the real prompt scrolled past the read window, only re-emitted last-prompt markers arrive, which never start a turn). A still-running delegation is now treated as proof the parent turn is live, so the notch stays "Thinking" instead of flickering to Idle between agent batches.
    • "Awaiting" (and "Limited") no longer disappear when you take your time. A session blocked on a question / plan approval goes quiet: it writes no new transcript lines until you answer, so once the wait passed the ~3 min active window Glint dropped it as stale and the notch stopped showing "Awaiting". Such blocked sessions are now kept live regardless of the window (still reaped after 24 h if truly abandoned), so a pending question stays visible until you reply.
    • "Awaiting" no longer lingers after you answer when Claude resumes by dispatching a subagent. The parent's own transcript stays silent while the agent works, so the notch used to sit on "Awaiting" for the whole delegated batch; a running subagent is now taken as proof the question was answered and the turn resumes.
    • A session no longer sticks on "Thinking" (showing "N agents") after it has actually finished. Subagent liveness now keeps a turn alive, but if an agent's own terminal line was never tailed its running flag could stay set forever. The parent's turn-terminal events (end_turn / interrupt / usage limit / measured duration) now finish every tracked subagent, since Claude can't end its turn while delegated work is still running, so the session falls idle when it should.
    • A Claude usage limit no longer hijacks the collapsed notch when another provider is live. Across providers the badge now surfaces active work (awaiting > thinking > limited > idle), so a rate-limited Claude session stops masking a Kiro/Codex session that's actually running. Within a provider a limit still ranks top.
    • A usage limit now clears itself once its reset time passes. Glint resolves the "Resets 5:40am (Asia/Colombo)" text to an absolute instant and drops the lockout at that time even if the session emitted no new lines, so "Session limit" no longer sticks around after you're back under the limit.
    sha256 597e2d1e34792497
  4. v0.1.8Beta2026-06-243.6 MB
    Added
    • The Usage tab now remembers which provider (Claude / Kiro / Codex) you last viewed and reopens on it, instead of always resetting to Claude. If that provider has since been disabled in Settings, it falls back to Claude.
    Fixed
    • The Usage tab no longer shows bogus extra meters ("Top skills", "Top subagents", "Top plugins"). Newer Claude Code CLIs append a "What's contributing to your limits usage?" breakdown below the plan meters, and Glint was parsing those lines as additional usage bars. It now stops at that section and reads only the real session/weekly meters.
    • A session no longer flips to a phantom "Thinking" when its editor (e.g. VS Code) is closed and reopened. Claude Code re-emits the last-prompt marker on every resume, after the turn already finished; Glint was reading that re-emitted marker as a brand-new turn. The marker now only sustains a turn a real prompt started, never creates one.
    • The expanded view now shows every session running concurrently in the same project, instead of collapsing them to one. Idle sessions still fold down to the most recent per project, so a /cleared predecessor stays hidden under its live successor.
    sha256 3fe962bc5c0aa65f
  5. v0.1.7Beta2026-06-183.6 MB
    Added
    • Codex support (opt-in, Settings → Other → Providers): with Codex enabled, live Codex CLI sessions from ~/.codex/sessions appear in the island, floating pill, and Dock bar alongside Claude and Kiro, with Codex's blue working sparkle, model, effort, context usage, active tool, recent tools, current plan, output tokens, and elapsed time. Reads Codex session JSONL locally; off by default.
    • The Usage tab's Codex side now mirrors the Claude tab: **Session (5h)** and **Weekly (7 day)** usage meters plus an account card (ChatGPT plan, email, account, Codex CLI version). Codex has no headless /usage command, so the meters read the rate_limits block Codex writes into every turn's session log
    • the same numbers the Codex TUI's /usage and the VS Code extension show, with no extra CLI run or token cost. Usage is shown as *used* (bars fill, red past 90%) to match Claude. A Settings toggle adds the 5h / weekly bars to the expanded island footer.
    Changed
    • Codex monitoring is lighter and stays flat under load. It used to rescan the whole ~/.codex/sessions history (one file per session, forever) and re-read + re-parse each active session's entire 512 KB tail on every 1s tick. Now it scans only the recent date-shards that can hold an active session, and reads each session **incrementally** - only the bytes appended since the last tick - so a streaming session parses a few new lines per second instead of half a megabyte, and the cost no longer scales with how many sessions are running at once. It also only republishes when the Codex sessions actually change, so an idle Codex no longer re-merges the session registry every second.
    • The expanded-island usage footer now lays each provider out on a single inline row (name + meters on one line) instead of a name line above its bars, so Claude, Kiro, and Codex stay compact together and their bars line up.
    Fixed
    • Codex sessions no longer fall idle in the middle of a turn. Codex emits several agent_message events within one turn (between tool batches); Glint was treating the first as the end of the turn, so the island dropped to "Idle" a few seconds after you submitted and never returned to "Thinking" until the next prompt. Only task_complete now ends a Codex turn.
    • The Kiro usage refresh button in the expanded footer now updates the meters immediately. It was fetching fresh data but the footer read a non-observable cache, so nothing repainted (most visible when idle). Codex now has its own footer refresh button, wired the same way.
    sha256 a8872ab68123eb73
  6. v0.1.6Beta2026-06-173.5 MB
    Changed
    • Kiro monitoring is much lighter on CPU and battery when several agents run at once. All session-file reading and parsing now happens off the main thread, parsed state is cached and only re-read when a file actually changes, and the live tool and turn detail are parsed only for the session you have expanded. Idle sessions no longer cost anything to keep on screen.
    Fixed
    • Kiro now shows "thinking" more reliably in the moment right after you submit a prompt, before the first token streams back.
    sha256 3cd6045bb7282dd5
  7. v0.1.5Beta2026-06-174.4 MB
    Added
    • Kiro support (opt-in, Settings → Other → Providers): with Kiro enabled, live kiro-cli sessions appear in the island, floating pill, and Dock bar alongside Claude, with a purple working sparkle, model, and context window. Sessions show the active tool, plan/todos, and the turn's credits and elapsed time, with thinking vs idle read live from the kiro-cli session log and lock file. Reads ~/.kiro/sessions/cli locally; off by default.
    • Kiro usage: with Kiro enabled, the Usage tab gains a Claude/Kiro selector; the Kiro side shows your plan's credit usage from kiro-cli (on-demand, no polling), plus an Account section (auth method, email, plan, and kiro-cli version).
    • Kiro credits in the expanded footer: an opt-in meter (Usage tab → Kiro) shows your Kiro credits on the notch, pill, and Dock bar, alongside the Claude meters.
    Changed
    • The panel header now shows a fixed tagline ("Agentic activity, right where you want it.") instead of echoing the active tab; your Claude plan name moved to a line atop the Usage card (mirroring the Kiro tab), so it's still one tap away.
    • The "show usage meters on expanded views" toggle moved out of Settings → Other into the Usage tab itself (one toggle per provider, next to that provider's meters). Your existing on/off choice for the Claude meters is preserved.
    sha256 a1cdeeb2d9975235
  8. v0.1.4Beta2026-06-143.4 MB
    Added
    • The expanded usage meters gain a refresh button at the start of the row - click it to refetch usage now; it shows the spinning sparkle while fetching (rapid taps coalesce into a single check). The API-spend view is live, so it doesn't need one.
    sha256 c3feb2425626bb9b
  9. v0.1.3Beta2026-06-143.4 MB
    Added
    • API/console accounts (billed per token, no Pro/Max plan) now get an estimated spend view instead of empty plan meters. The Usage tab shows estimated cost and a token breakdown (input / output / cache) for the sessions Glint is tracking, and the expanded usage meters show a compact "~$X - N tok". The cost reuses the existing cache-aware estimate at public API rates - clearly labelled an estimate, computed locally with no API key and no new network calls.
    Changed
    • When an update is available, the Other tab's button becomes "Get the update" (opening the downloads page) instead of showing it next to "Check for Updates".
    • The expanded usage meters show a spinning sparkle while a usage refresh is in flight, so a stale-then-updating read is obvious.
    sha256 f2c523c542861e00
  10. v0.1.2Beta2026-06-143.3 MB
    Added
    • Check for Updates (Settings → Other): a button that, only when clicked, asks whether a newer Glint is out. It shows a spinning sparkle while checking, then either "up to date" or the new version with a button to the downloads page. No background polling - it checks once per click against the published release.
    • Usage meters (opt-in, Settings → Other): the expanded view can show your Claude Code session and weekly usage as two small bars next to "Quit Glint". They read the same cached usage as the Usage tab (no background polling) and refresh only when you expand, at most once every 120 seconds. Off by default; when off they read and fetch nothing.
    sha256 fa5cf16d2d8b526f
  11. v0.1.1Beta2026-06-143.3 MB
    Added
    • Scroll to expand (opt-in, Settings → Other): hover a surface and scroll on the trackpad to open or close it. The notch opens scrolling down, the Dock bar scrolling up, and the floating pill follows which half of the screen it's on. Scroll the other way over the island's top strip to collapse; the expanded panel's own list keeps its scrolling.
    sha256 b4c959a9b7661487
  12. v0.1.0Beta2026-06-143.3 MB
    Added
    • The dock-side bar has a width setting: Automatic (fills the whole empty space beside the Dock, as before) or Custom, where a slider sets a fixed width (120-1400 pt) that hugs the Dock and clamps to the space available. Handy when you only have a few Dock apps and the bar would otherwise span the whole screen.
    Changed
    • The dock-side bar now expands to a comfortable panel that hugs the Dock and grows outward, instead of being locked to the collapsed bar's width. A thin custom bar no longer squeezes the expanded view; a very wide gap no longer stretches it edge-to-edge.
    • A narrow dock-side bar sheds detail to stay legible while working: a roomy bar shows the full activity line plus elapsed and token metrics; a medium bar shows just the status word (Thinking / Awaiting); a sliver shows only the status glyph.
    sha256 e0e9c89d3740a7ed
  13. v0.0.10Beta2026-06-123.3 MB
    Fixed
    • Claude Code 2.1.162 and older run /usage headlessly but print only the subscription preamble with no meter rows (verified against archived CLI builds; rows appear by 2.1.169). The Usage tab now recognizes that shape and says to update the CLI instead of echoing the preamble.
    • A fully reset account prints the same no-meters /usage output on CURRENT CLIs, which falsely triggered the "update your CLI" message. The panel now uses the installed CLI version to tell the two apart: a current CLI shows calm zeroed meters with "All usage windows have reset - the meters fill back in after your next Claude Code turn", and a freshly-reset row that prints without a reset clause reads "Nothing used yet" instead of "Reset time unavailable".
    sha256 a44bcbeb3c3737fd
  14. v0.0.9Beta2026-06-113.3 MB
    Added
    • Every surface (notch island, floating pill, dock-side bar) is now a full entry point: right-click any of them for Open Glint / Settings / Quit, and the expanded view gains a gear that opens the panel - anchored below the island, so both work even when macOS hides Glint's menu icon in an overflowing bar. "Settings..." lands on that surface's own tab.
    • While working, the notch island stops short of real menu bar icons instead of covering them, so the icons stay clickable - including Glint's own. Status-item positions are read without any permission; the frontmost app's menu extent is also respected when Accessibility is granted (the Dock Side permission). Toggle: Settings -> Notch -> "Keep menu bar icons clickable" (on by default).
    • The Account card shows the installed Claude Code version, so a tester's setup is diagnosable from a screenshot.
    • A "Quit Glint" link sits under the version in the menu-bar panel header, so quitting no longer requires expanding an island.
    Changed
    • First-run defaults: the floating pill is the surface that starts enabled (it works on every Mac, notched or not); the notch island and dock-side bar are opt-in. Existing installs keep their choices.
    • Settings reorganized: the pill's per-display jump buttons are now a "Move to display" dropdown matching the notch tab's display picker. The "When idle" option is hidden - every surface uses the quiet default (a small dot beside a real notch, a compact pill elsewhere).
    • The accent color option is hidden for now - every surface uses the brand orange. Previously customized colors revert to the default (the stored choice is kept for when the option returns).
    • On screens without a physical notch, the notch island no longer reserves a fake 200pt center gap: it idles as a compact pill and grows with its text while working (long tool lines stretch it, with the same springy animation), instead of sprawling as a wide half-empty bar.
    • Beside a real MacBook notch, the working island is quiet by default: just the status glyph (spinning sparkle / pulsing dot / warning triangle) next to the camera. Hovering stretches it open to reveal the status word, tool line, and time + token metrics - Dynamic-Island style.
    Fixed
    • Older Claude Code versions gate /usage behind an interactive terminal ("/usage isn't available in this environment"); the panel now explains the fix ("Update it (claude update), then refresh") instead of echoing the raw CLI message. Headless /usage support is version-dependent, not plan or seat-tier dependent - Team Pro/Max seats both report fine on a current CLI.
    sha256 d45666947d12c671
  15. v0.0.8Beta2026-06-113.2 MB
    Fixed
    • The Usage tab handles non-Max accounts properly. Team seats (verified live) report the same "NN% used" meters minus the Sonnet row, which now renders as two meters instead of an error; the percent matcher accepts more phrasings; older CLIs that reject --no-session-persistence get a legacy retry; and if nothing parses, the panel shows what /usage actually said so the cause is visible at a glance.
    sha256 ccdeb68625c07992
  16. v0.0.7Beta2026-06-113.2 MB
    Added
    • New brand logo everywhere: the app icon (Finder/Dock/Launchpad) and the in-app marks (panel, settings, paywall headers) now use the orange-gradient sparkle artwork instead of the drawn accent square.
    • The Usage tab's refresh button shows the island's spinning sparkle (in white) while a fetch is in flight, and ignores clicks until it finishes.
    Changed
    • The menu-bar panel is half the size (400×500, was 520×690) with a tighter header, compact tab bar (tabs renamed "Pill" and "Dock", stretched evenly across the full width), and the Usage section above Account.
    • The panel background is now the system's Liquid Glass on macOS 26 (frosted material on older macOS) instead of an opaque window color.
    Fixed
    • "Manage in Polar portal…" opens Polar's actual customer portal (email + login code, where buyers can free device slots) instead of the checkout page.
    sha256 52653d12fa4256a6
  17. v0.0.6Beta2026-06-110.8 MB
    Added
    • The session row now shows the reasoning effort in use (e.g. "Ultracode Xhigh") between the model and status. The session's own /effort override wins; otherwise the global level from ~/.claude/settings.json, with ultracode detected from the transcript's own mode markers - no polling, no extra processes.
    Fixed
    • The context gauge no longer pegs at 100% on 1M-context sessions. The window size was hardcoded to 200k; sessions on the 1M-context beta (whose transcripts don't announce it) are now detected from their token totals, and the percentage matches Claude Code's own "until auto-compact" gauge instead of raw window fullness.
    sha256 7f82243389da1795
  18. v0.0.5Beta2026-06-110.8 MB
    Changed
    • The expanded island shows context-window fullness as a compact circular gauge (the percentage sits inside the ring, which turns red past 90%) instead of the "· 85% ctx" text in the session row's meta line.
    Fixed
    • Long-running Bash commands (builds, tests, installs) no longer flip the island to "Awaiting" - and no longer ring the needs-you sound. Only instant-completing gated tools (Write/Edit) treat a stalled approval as "blocked on you"; a running command now just shows its live tool verb while the session stays Thinking.
    • Glint no longer shows Idle while Claude is actually working after a queued message: prompts typed while Claude was busy land in the transcript as array-content user lines, which weren't recognized as prompts - the island sat Idle from the queued prompt through the model's silent thinking phase until the first response line. Queued and image prompts now start the turn immediately.
    • Brand-new session transcripts (e.g. right after /clear) are picked up instantly via per-project folder watchers instead of waiting for the slow rescan.
    sha256 1e2499c45130d2ba
  19. v0.0.4Beta2026-06-110.8 MB
    Changed
    • Massive performance overhaul - Glint now idles at ~0% CPU (was ~32%) and cold start is instant even with hundreds of MB of transcripts:
    • Launch backfill parses only the tail of recently-active transcripts, skips stale ones entirely, and runs off the main thread - the island appears immediately instead of beachballing through a full history scan.
    • Only actively-written transcripts are watched and polled (one or two files instead of every session ever recorded); everything else is rescanned on a slow cadence and picked up the moment it resumes.
    • The UI republishes at most once a second and only when something actually changed, so an idle Glint does zero rendering work.
    • Sessions idle for over a day are pruned from memory (usage-limit lockouts are kept so weekly reset countdowns survive).
    • The Usage tab no longer runs the Claude CLI on a background loop - it refreshes when the panel opens (with a short cache), and the CLI runner can no longer deadlock or hang forever (drained pipes + watchdog timeout).
    • The dock-side bar polls the Dock a couple of times a second, bursting to display rate only while the Dock is actually animating (was a constant 60 Hz timer even when the surface was disabled).
    Fixed
    • Huge pasted prompts are no longer retained in full per session, finished subagent records are cleaned up, and launch-at-login no longer does a blocking system call during startup.
    sha256 f046ded558542664
  20. v0.0.3Beta2026-06-110.7 MB
    Changed
    • Clicking the menu-bar icon now opens a larger tabbed panel with Claude account usage, Settings, and License activation in-place instead of showing the old dropdown menu or opening separate settings/activation windows.
    • The menu-bar Usage tab now shells out to Claude Code's own noninteractive /usage command to show session, weekly, and Sonnet percentages, with a usage card matching the account section, muted rails, accent-filled bars, and a taller panel that avoids scrolling for the full usage section. The extra Local Signals diagnostic card was removed.
    • Claude usage is warmed in the background and cached across menu-bar panel opens, so the Usage tab shows stable progress rows immediately and animates the bars when refreshed data arrives.
    Fixed
    • Claude account status now inherits the app environment when running claude auth status --json, so an authenticated CLI is no longer reported as unavailable just because Glint launched it with a stripped environment.
    • Claude Code session/weekly usage-limit errors now surface as a limit-hit status with reset text, and the island stops the running timer instead of staying stuck on Thinking.
    sha256 41aa63bc3e6e371d
  21. v0.0.2Beta2026-06-100.7 MB
    Added
    • One-time lifetime license via Polar ($6.99 launch offer, up to 2 Macs): buy on the website, paste the key into the paywall or the new Settings → License tab to activate. Validates with Polar on launch and daily, with a 24-hour offline grace period.
    • Hard paywall: the island surfaces and session monitor only run once a license is activated. The menu bar gains "Activate Glint…" while locked.
    Changed
    • Privacy note: the license check is the only network call Glint makes - session data still never leaves your Mac.
    sha256 3ec957c959deaeb8
  22. v0.0.1Beta2026-06-090.8 MB
    Added
    • Initial beta. Glint surfaces what your Claude Code sessions are doing, read live from ~/.claude/projects/** - nothing leaves your Mac.
    • Live status - Idle / Thinking / Awaiting (blocked on you), with the active tool verb.
    • Per-turn metrics - output tokens, cost, and elapsed time, matched to Claude Code's status line and reset every prompt.
    • Plan, Subagents, and Context window - the TodoWrite checklist, how many agents are out, and how full the 200k window is.
    • Three surfaces, each toggled independently: a notch island, a draggable floating pill, and a dock-side bar.
    • Optional per-event notification sounds on status changes (all off by default).
    • Distributed as an ad-hoc-signed .dmg. Apple Silicon (arm64), macOS 14 (Sonoma) or later.
    sha256 818bc3487189d105