Every macOS agent-notch app compared (2026)

The field at a glance

Vibe Island's star count is for its public community repository, and its date is its latest release, from its changelog; the app itself is closed source, as are agentnotch.app, Claude Pulse and NotchClaude, which have no public repository to count. TokenTracker, claude-control and both apps called Claude Status are not notch apps; they are here because they watch the same agents.

Each app, and where it beats Notchmeter

Vibe Notch (formerly Claude Island)

farouqaldori/vibe-notch. Claude Code sessions in a Dynamic Island-style overlay, with permission approvals and the full chat history rendered in the notch.

Where it beats Notchmeter: the largest community on this page, and two other apps here, MioIsland and Buddi, started from its code. No setup at all. It shows a session's conversation, which Notchmeter deliberately does not.

Worth knowing: its README says it installs its hooks automatically on first launch and collects anonymous usage events with Mixpanel (app launched, session started). Its README calls it actively maintained; the last push was on 2026-04-20.

CodeIsland

wxtsky/CodeIsland. "Your AI coding agents, live in the MacBook notch." Live status, the current tool and the latest reply for every session of more than 30 AI coding tools, with approvals, answers to questions, and a jump to the exact terminal tab, IDE window or pane.

Where it beats Notchmeter: breadth and reach. It covers far more tools, answers approvals and questions for many more of them, and shows sessions on SSH hosts beside local ones. It has iPhone and Apple Watch companions (the iPhone's Dynamic Island, Lock Screen and StandBy), an ESP32 desk display, pushes to Bark, ntfy, Slack, Telegram and others while you are away, and webhooks for your own automations. Each tool gets a pixel-art mascot, and there are quiet hours and rules that silence a directory.

Worth knowing: its README says it writes its hooks into each detected tool's config at launch and repairs them if they drift, each one switchable off under Settings › Hooks. Its usage view is Claude's stats, with the five-hour and weekly limits opt-in. MIT, macOS 14 or later, and seven languages.

Open Island

Octane0411/open-vibe-island. Session status, permission approvals and a jump back to the right terminal for 13 agents, in the notch or, on a display without one, a bar at the top centre. Its README says: "Think of it as an open-source Vibe Island".

Where it beats Notchmeter: more agents, Grok Build, Pi, Qwen Code, Qoder, Factory and CodeBuddy among them, and more places to jump back to: Zellij, cmux and Kaku panes, and the workspace in JetBrains IDEs, Windsurf, Trae and Zed. It answers OpenCode's permission requests and questions, which Notchmeter leaves to OpenCode's own prompt.

Worth knowing: GPL-3.0 and free. Hooks are installed and removed from its Settings window; its README says the OpenCode plugin is installed automatically on first launch, and that its Claude status line is opt-in and will not replace one you already have. Its usage view covers Claude and Codex. macOS 14 or later, in English and Simplified Chinese.

TokenTracker

xiufengsun/TokenTracker. A local token and cost dashboard for 43 AI coding tools by its README's count (its repository's description says 31), with a macOS menu-bar app, four desktop widgets, a desktop pet and fifteen achievement tracks. It is not a notch app.

Where it beats Notchmeter: breadth, and reach beyond the Mac. It counts tokens for far more tools than Notchmeter meters, shows quota windows for seventeen providers, and runs on Linux and Windows as well, as a command line tool and a tray app. It draws your activity as a GitHub-style heatmap, and its widgets put usage and limits on the desktop.

Worth knowing: usage, not sessions. Its first run installs a SessionEnd hook or a notify entry in each detected tool's own config, and tokentracker uninstall removes them all. A global leaderboard and cloud sync across machines are optional, and the leaderboard needs you to sign in. The menu-bar app shows the dashboard in a web view.

Ping Island

erha19/ping-island, with a site at erha19.github.io/ping-island. A menu-bar app that opens from the notch when a coding session needs you, for the seventeen clients its README lists, from Claude Code and Codex to Hermes Agent, OpenClaw and GitHub Copilot. Its README places it as an open-source alternative to Vibe Island.

Where it beats Notchmeter: SSH hosts are built in: it sets up a bridge on the remote machine and points the remote hooks back at your Mac. It answers approvals and questions for more clients, can auto-approve a Claude Code session you choose, and lets you drag its mascot out of the notch as a floating companion. Three visual themes, and sound packs you can import.

Worth knowing: Apache-2.0 and free, macOS 14 or later. Its privacy policy describes optional anonymous telemetry: the first run offers it with the consent box already ticked, nothing is sent until you confirm, and it never includes prompts, code or paths.

Notchi

sk-ruban/notchi. A notch companion for Claude Code and Codex: a pixel sprite per session that reacts to thinking, working, permission requests, compaction and errors, with session time, usage quota and 30 days of cost and tokens a click away.

Where it beats Notchmeter: personality, and a real community. A sprite per session says what each agent is doing at a glance, in a way a ring does not, and it is actively maintained.

Worth knowing: it installs its hooks on first launch. An optional mood feature sends your prompts to Anthropic or OpenAI for sentiment analysis, with an API key you paste in. It needs a MacBook with a notch.

MioIsland

MioMioOS/MioIsland. A Claude Code notch app, forked from Claude Island, with an iPhone companion called Code Light, status, session lists, approvals, buddies and 8-bit sound.

Where it beats Notchmeter: your sessions on your phone, with two-way sync and remote launch. Notchmeter has no phone app.

Worth knowing: its README states a non-commercial licence (CC BY-NC 4.0), which GitHub does not recognise as open source. The README is mostly images, in Chinese.

CodexIsland

ericjypark/codex-island. "Your AI usage limits, living in your notch": Claude Code's and Codex's 5-hour and weekly windows, local-log cost estimates, and a year of usage history.

Where it beats Notchmeter: it is the app closest to Notchmeter's own pitch, and it is moving fast (0.3.0 shipped on 2026-10-01). Its usage history is a year-at-a-glance calendar kept in its own database, it offers five chart styles, and it builds for macOS 13. It also makes a shareable usage card and estimates OpenCode's cost, as Notchmeter has since 0.9.0.

Worth knowing: Claude and Codex only, and no sessions. Its README says the app is unsigned by Apple and its Homebrew cask removes the quarantine attribute. It reads the credentials Claude Code and Codex already keep on the Mac, and never refreshes them.

agent-notch

realfishsam/agent-notch. "No hooks, no APIs, no accounts." A session is an agent process attached to a terminal, polled every three seconds; a mascot walks beside the notch while it works.

Where it beats Notchmeter: it never writes anything and needs no setup, and it is small enough to build from one Swift file. Its README is unusually honest about the cost of that design.

Worth knowing: Claude Code and Codex only, terminal sessions only, and no waiting detection. In its own words, the mascot "keeps walking for up to ~33 s … after a turn actually ends".

AgentNotch (AppGram)

AppGram/agentnotch. Real-time tool calls, tokens and estimated cost for Claude Code and Codex, received as OpenTelemetry on port 4318.

Where it beats Notchmeter: it sees every tool call as it happens, from the agents' own telemetry, without a hook in any settings file; you point the agent at it yourself.

Worth knowing: this is not the same product as agentnotch.app below, despite the name. Its README says MIT, but GitHub finds no licence file, and there has been no push since January.

Vibe Island

vibeisland.app. A notch panel for a very long list of agents: live status, approvals, answers to questions, plan review, and a jump back to the exact terminal tab, split pane or IDE window.

Where it beats Notchmeter: breadth, clearly. Its site lists 30 agents (its repository description says 25) and 20 terminals and IDEs. SSH remote monitoring is built in, where Notchmeter's needs a tunnel and a hook you set up yourself. It reviews full Markdown plans, and its 8-bit sounds can be replaced with your own packs. Its usage view covers Claude, Codex, Kimi, GLM and DeepSeek.

Worth knowing: closed source; $19.99 for one Mac with a year of updates, and a free trial. Its README says detected CLIs are set up automatically on first launch, and that it needs macOS 14 or later. Its changelog shows a task checklist on each session card and rules that silence noisy sessions, close to Notchmeter's task lists and muted projects. It publishes its own comparison of the field, which is worth reading beside this one.

claude-control

sverrirsig/claude-control. A desktop dashboard for many Claude Code sessions at once: each one's status, branch, changed files, pull request and CI state, and its last message, with permission prompts approved or rejected from the dashboard. It is not a notch app: it is a window.

Where it beats Notchmeter: it runs the work as well as watching it. It starts new sessions, in a new git worktree if you want one, sends /create-pr to an idle session, shows a pull request's checks, review and conflicts, and removes the worktree and branch when you are done, with keyboard shortcuts for the common actions.

Worth knowing: Claude Code only, and no usage limits. Its README says it installs its hooks into ~/.claude/settings.json automatically on first launch. It is built with Electron and Next.js.

Buddi

talkvalue/Buddi. A home in the notch for Claude Code's buddy, built on boring.notch and Claude Island, with Claude's session and weekly usage beside it.

Where it beats Notchmeter: it is a whole notch utility, with a music player, calendar, file shelf and volume and brightness overlays beside the agent. It has eighteen buddy species, a chat view that can send messages back to Claude Code, and tmux and cmux support.

Worth knowing: Claude Code only, and it installs its hooks on first launch.

Claude Stats

1pitaph/claude-stats. A menu-bar stats app for AI coding work, with a notch island among many other modules.

Where it beats Notchmeter: scope. Dashboards, Git and GitHub views, daily reports, an iPhone and iPad companion through iCloud, an embedded terminal and a network debugger, and a Lite build for people who want less.

Worth knowing: it parses Claude Code and Codex logs today; the notch is one module, not the product. Its current releases are for Apple silicon only, and its iOS companion is built from source rather than published.

Claude Status (gmr)

gmr/claude-status. A menu-bar monitor for every Claude Code session: one indicator for the state across them all (active, waiting, compacting or idle), a list to jump from, and desktop widgets. It is not a notch app.

Where it beats Notchmeter: desktop widgets, which Notchmeter does not have, and a jump to the exact window, tab or pane in hosts where Notchmeter only brings the app forward, JetBrains IDEs, Zed and Xcode among them. It also records how long each session spent in each state and how many ran at once through the day.

Worth knowing: Claude Code only, and no usage limits. It asks on first launch before installing its Claude Code plugin hook. It needs macOS 26.2 or later. Not the same app as Claude Status (runlocal.dev) below.

ClaudeNotch

rawsun007/claude-notch, with a site at rawsun007.github.io/claude-notch. Claude Code's permission prompts, questions and notifications in the notch, with an edit's diff inside the card, and Codex sessions shown in beta.

Where it beats Notchmeter: depth on Claude Code's prompts. Risky commands such as rm -rf, sudo and a force-push need Touch ID or a press-and-hold; a denial can carry a note telling Claude what to do instead; a finished task has a Reply button; and dropping a folder on the notch starts a session there. Spending caps per session and per day, where Notchmeter's budgets are weekly or monthly, a claudenotch:// URL scheme and AppleScript, and a pet that lives in the notch.

Worth knowing: Claude Code fully and Codex in beta, where the notch reports a session but does not answer it. Free for personal and noncommercial use under its own licence, which GitHub does not recognise as open source. Its hooks go in when you click Setup, after a backup of settings.json. macOS 13 or later, in English and nine other languages.

QuotaBar

QuotaBar/QuotaBar, with a site at quota.bar. "Every AI coding limit, at a glance": each provider's quota windows, resets and estimated spend, in the menu bar, either side of the notch, in a dock at the screen's edge or as cards on the desktop.

Where it beats Notchmeter: providers, by a long way: 21 by its site's count, among them Grok, z.ai, MiniMax, DeepSeek, Qwen, OpenRouter, Windsurf and Kiro beside the ones Notchmeter reads, several of them marked experimental. Desktop cards in seven styles, eleven menu-bar glyphs, the providers' own status pages, and an iPhone app with widgets in the same repository, fed through iCloud.

Worth knowing: limits and spend only, no sessions. Its README says twenty-two providers where its site counts 21. Its update check goes to quota.bar, which, by its README, keeps a salted hash of your IP address to count active installs and no usage data. Quota Run, an opt-in leaderboard, needs an account; the app works fully without one.

Claude Status (runlocal.dev)

bcollard/claude-status-macos-menu-bar, with a site at claudestatus.runlocal.dev. Claude Code's plan limits, today's spend and a per-model breakdown in one menu-bar item. It is not a notch app.

Where it beats Notchmeter: it is small and does one thing, in about 1,500 lines of Swift by its README's count, and it runs on macOS 14, which Notchmeter no longer supports.

Worth knowing: Claude Code only, and no sessions. Apple silicon only for now; its README says a universal build is coming. Its call to Anthropic's usage endpoint can be turned off, leaving the local logs. Not the gmr app above.

Headroom

PeytonNowlin/Headroom. "A quota island for your MacBook notch": limits for Claude, Codex, Grok, Cursor and OpenCode Go, as gauges beside the notch, rings on hover, and every window with its reset on a click.

Where it beats Notchmeter: restraint. It draws nothing at all while every main agent has room, and a provider you mark as a side one never brings the island on screen by itself. It reads the Grok CLI's quota, which Notchmeter does not.

Worth knowing: limits only: no sessions, and since it stopped reading session logs, no spend estimate. It needs macOS 26 on Apple silicon, and its README says it is ad-hoc signed rather than notarised, so the first launch needs Open Anyway.

AgentNotch (agentnotch.app)

agentnotch.app. "Every coding agent. One place to look." Sessions first, usage second, for Claude Code, Cursor, Codex, Kimi, OpenCode, Cowork, Antigravity, Pi and LM Studio, per its site, with plan usage for Claude, Codex, Cursor, Kimi and OpenCode Go.

Where it beats Notchmeter: it feels finished, since 2.1.7 it shows Cursor's cloud agents beside local ones, and its feedback from inside the app is anonymous, where Notchmeter's is filed under your own GitHub account or sent from your own address, because Notchmeter has no server to be anonymous through. The rest of what it had first, Notchmeter has matched since 0.9.0: a choice of panel colours and of how much desktop shows through, and a light Paper theme; distinct sounds for completions, questions, approvals, ready plans and other requests; a locally computed meter for the OpenCode Go plan, which publishes no usage endpoint; and a switch per display.

Worth knowing: closed source; $9.99 once, with a three-day trial that needs no card and a 14-day refund. Its site does not say how it reads sessions; its changelog records that its Live Prompts "can no longer overwrite your agent's configuration, and every write leaves a backup". Not the AppGram project above.

Claude Pulse

claudepulse.app. Every Claude Code session on the Mac in the notch, with tool calls approved or denied there, and the five-hour window, the weekly limit and each session's context fill beside them.

Where it beats Notchmeter: approval in batches, with rm -rf, DROP TABLE and a force-push always left for you to confirm, and a notch that bounces instead of opening when you are already looking at that session.

Worth knowing: closed source and free, with no account. Claude Code only. Its changelog says its hooks install automatically, and its latest entry is v1.0, from April 2026. macOS 15 or later, on a MacBook with a notch and Apple silicon.

NotchClaude

notchclaude.com. "Claude usage in your notch": Claude's five-hour and weekly windows with their resets, which Claude account is active, and spend once pay-per-use starts.

Where it beats Notchmeter: simplicity. It is one meter and nothing else, read from the Claude Code login already on the Mac.

Worth knowing: closed source and free, downloaded from its site. Claude Code only, and no sessions. It needs a login made in Terminal with claude login, and it polls every 60 seconds. Its privacy policy says it has no server, no analytics and no telemetry. Its site names no version or date.

Notchmeter

Amir-Hackett/notchmeter. Usage meters for Claude Code, Codex, Cursor, Gemini CLI, Antigravity, GitHub Copilot, Kimi Code and OpenCode, with sessions, answering from the notch, and advice about what to do next.

Each session's row carries its task list, crossed off as it goes, for every assistant with a hook or plugin; Cursor's hooks never report the list, so it is read from the conversation's own transcript, and since 0.9.15 a Cursor plan is on its row before it is built, with Build. Where the assistant reports them, the row also names the model and shows Compacting while the session summarises its context. When a session waits for you or finishes a turn, the news reads across the notch, and a long name scrolls through once. Claude Code's permission requests and multiple-choice questions are answered from the notch, and so are Codex's and GitHub Copilot CLI's permission requests; Cursor's shell commands and its own Run and switch-mode cards are too, behind two switches that are off until you turn them on. A project can be muted for possible waits, a Cursor turn gone quiet or Claude Code's idle reminder, while its permission prompts and questions still reach you, and a Cursor nudge that turns out to be wrong waits twice as long the next time. Kimi Code reports no wait at all, so its sessions never show one.

Where it is behind: one star, so very few people have looked at it. Fewer agents than Vibe Island, CodeIsland, Ping Island or Open Island, and a different set from agentnotch.app's, without its Pi, LM Studio or Cursor cloud agents; far fewer usage providers than QuotaBar or TokenTracker. Approvals from the notch for four assistants, Cursor's behind a switch that is off until you turn it on, where CodeIsland and Ping Island answer more. No mascot, no phone app, no desktop widgets, and remote sessions need a tunnel you set up yourself, where CodeIsland, Ping Island and Vibe Island have SSH built in. It needs macOS 15, where several apps here run on 14 or earlier.

Worth knowing: the first hook for each assistant is added only when you press a button, after a backup (what works with no hook). After that, Repair an out-of-date hook at launch, on by default, rewrites Notchmeter's own entries and adds the ones a new version needs, after the same backup; that is how 0.9.13's new entries arrive. With it off under Settings › Integrations › Hooks, nothing is written without a click.

What it does that we did not find in the others' own documentation:

• A published accuracy document: every rule behind every figure, its primary source and date, and where it is known to differ from a bill, pinned by golden tests to nine decimal places.
• Every window labelled with where its figure came from, and a tool that cannot derive spend shown as missing, never as $0.
• A week or a month projected at how far it rose over the last day, nights and breaks included, so a week spent on its first night does not read as running out; the five-hour session at the pace of the hours you worked in it, with a run-out range; and peak hours kept apart. QuotaBar, Headroom, ClaudeNotch and Claude Stats forecast too: QuotaBar from the window's use so far, Headroom from up to an hour of readings, and ClaudeNotch at the rate you are spending.
• Advice worded as instructions, including routing work across vendors.
• Prompt-cache miss diagnostics, priced. ClaudeNotch shows what the cache saved.
• An MCP server and a Claude Code plugin, beside a command line tool, a local API and CSV or JSON export. QuotaBar also has a local API and JSON from the command line, TokenTracker a command line tool, and ClaudeNotch a URL scheme and AppleScript.
• Eleven languages, a VoiceOver label on every readout, and status never carried by colour alone. ClaudeNotch, in ten languages, announces its cards to VoiceOver, and Headroom's README also says colour is never its only signal.

The one from Anthropic

If you work only in Claude's desktop app, you may not need any of these. Its documentation says it "sends an OS notification when a Code session finishes a task and you aren't currently viewing that session", lets you "filter sessions by status, project, or environment, and … group sessions by project", and shows your context and plan usage on a ring next to the model picker. It is first-party, and it costs nothing extra. What it does not do is meter Cursor, Copilot, Codex or Gemini, and that multi-vendor view is the ground most apps on this page stand on.

Beyond the notch

Beyond the four apps above that do not use the notch, a larger group tracks usage from the menu bar, among them ClaudeBar, coding-usage-bar, ClaudeUsageBar, Claude-Usage-Tracker and claude-code-monitor, and another browses past sessions, such as Claudoscope, Agent Sessions and so-agentbar. Of the notch apps Vibe Island's own comparison names, one is not reviewed above: Notchcode (31 stars), which also runs on Windows; it is not the paid NotchCode at notchcode.dev, which neither comparison covers. None of these was reviewed for this page. Notchmeter's own pace projection and cost rules started from OpenUsage, which ports ccusage's transcript semantics, and both deserve a look.