Know your limits.
Codex and Claude Code usage with reset times, warnings at 80% and 95%, and a pace forecast.
5 hours · 61%at this pace: 100% at 16:40
Your Codex and Claude limits, right in the notch.
Follow every agent session live, get pinged when one finishes or needs you, and dictate with Space. Free and open source for Mac, Windows and Linux.
100% open source · MIT licence · read the code
Or install from Terminal — no security warning:
git clone https://github.com/pnpmx/AgentNotch.git && cd AgentNotch && scripts/install-local.sh
Paste it in Terminal and press Enter. If macOS offers to install developer tools, accept and run it again.
macOS 26+ · Windows 10 & 11 · Linux (X11 and Wayland) · free and MIT licensed
Built by Pietro, a solo developer
WHAT IT DOES
Your agents,
one glance away.
Limits, live sessions, alerts and voice. In the notch on Mac, on the screen edge on Windows and Linux.
Every session, live.
All your Claude Code and Codex sessions at once, across terminal tabs: what each one is doing right now.
my-app running npm testapi waiting for youdocs done · $0.42 · 4 min
Never miss a finish.
Orbiting lights show which agents are alive: orange for Claude, mint for Codex. Double-click a session to jump to its terminal tab.
Claude alive Codex aliveClaude finished · my-app↗ jump to Terminal tab
Pick model and effort.
Set the default model and effort for new Claude Code and Codex sessions in two clicks.
Claude Code Opus · HighCodex GPT-6.1 · Medium
Think it. Say it.
Hold Space to dictate into any app, transcribed on your computer, with your own vocabulary and no filler words.
Listening…“Add tests for the login flow”
Drop, hand off, share.
Drop files on the notch to paste their path, continue a task in the other agent, and share your week.
⇣ screenshot.png → path pastedContinue in CodexYour week: 23 tasks · +4,210 lines
At home on each system.
Swift in the Mac’s notch. A slim edge tab on Windows and Linux. Same quiet idea on both.
Your voice, locally.
Apple SpeechAnalyzer on Mac and whisper.cpp on Windows run on-device. No hosted speech API key.
Open from the start.
MIT-licensed app source. Read it, change it, help make it better.
NOW ON WINDOWS & LINUX
No notch?
No problem.
A slim tab that lives on the edge of your screen.
Any edge, any screen.
Drag it to the left, right, top or bottom of any monitor. Let go and it snaps to the closest edge, and remembers where you like it.
Stays out of your focus.
Click it to open the full panel without taking focus away from the app you’re typing in.
Talk, locally.
Hold Ctrl Shift Space, speak, release. Whisper runs on your PC; the text lands where your cursor was.
Experimental · Windows 10 & 11 installer (unsigned, so SmartScreen will ask first) · Linux AppImage and .deb, including Wayland. Linux notes
OPEN SOURCE
Open source.
No catch.
Every line of AgentNotch is public on GitHub under the MIT licence. Free today, free tomorrow, no account, no subscription.
Read every line
Nothing hidden: check exactly what runs on your computer. Mac code · Windows & Linux code.
Yours to change
The MIT licence lets you use, modify and share it, even in your own products. Just keep the credit.
Private by design
No account, no analytics, no tracking. Your voice is transcribed on your device and never leaves it.
SEE IT IN MOTION
34 seconds.
You’ll get the idea.
Instrumental audio, no narration. Shows usage limits, hold-to-talk dictation and on-device speech with sample data.
LATEST UPDATES
What’s new.
THIS IS JUST THE START
Small tools.
Good little differences.
We’re making more tools for the moments between the work.
Less friction. Less noise. More room for what you actually want to do. AgentNotch is the first chapter—not the whole story.
Follow what we’re building ↗
BEHIND AGENTNOTCH
Hi, I’m Pietro.
Business by training, builder by habit.
I studied Business Administration (BBA) and then a Master’s in AI Applied to Business. That mix shapes how I work: I start from the business problem, then build the software that solves it.
Today I’m a full-stack developer with more than ten projects shipped end to end, from the first conversation to production, support and the next iteration. AgentNotch is my first open-source project.
10+projects shipped end to end
BBA + MScAI Applied to Business
3operating systems for AgentNotch
What I build
Operations platforms
Inventory, HR, time-tracking and patient workflows for a multi-site healthcare clinic, moved from spreadsheets to a SQL backend with automated email notifications.
Voice and chat AI agents
Phone receptionists that check availability, book appointments and look up clients in the CRM, plus WhatsApp assistants and outbound campaigns.
Integrations and data
CRM sync services, purchase forecasting and reporting that connect the tools a business already relies on.
Websites and platforms
Websites for restaurants and local businesses, a learning platform, and a digital-procedures platform for municipal governments.
Quant research
A research-first futures backtesting and validation system, and crypto arbitrage tooling.
Developer tools
AgentNotch on macOS, Windows and Linux, and Forge, a local studio for building AI applications.
TypeScriptReactNode.jsNestJSPostgreSQLPrismaSwiftRustTauriVoice AILLM APIsDockerVercel
HELP THE NEXT LITTLE IDEA HAPPEN
Like what we’re making?
Support independent development and the tools that come next.
No
pressure. Using it, sharing it, or opening an issue helps too.
Optional support. The app’s source stays MIT-licensed.
A few good questions.
What do I need to run AgentNotch?
Mac: an Apple Silicon Mac running macOS 26+, and Swift 6.2+ to build. Accessibility, microphone and Speech Recognition permissions are required for dictation. Windows / Linux: Windows 10 or 11 (x64), or a 64-bit Linux desktop (X11 or Wayland), and a microphone. The first dictation downloads a speech model once (about 150 MB). On both, sign in to Codex CLI separately for Codex usage and connect Claude Code through its optional status-line integration.
Is there a one-click app download?
Yes. Mac gets a disk image (.dmg): open it and drag
AgentNotch to Applications. Windows gets an installer and Linux an
AppImage or .deb. They are built by the project’s public
CI but not yet signed, so the first time macOS asks you to confirm in
System Settings → Privacy & Security → Open Anyway, and Windows
SmartScreen shows More info → Run anyway. Prefer to build it
yourself? Follow the
build instructions.
Does my voice leave my computer?
No. Speech recognition runs on-device: Apple’s models on Mac, whisper.cpp on Windows and Linux. Initial language assets may need to download. Codex and Claude usage integrations may contact their services; anything you choose to submit to an AI tool follows that tool’s own policies.
Which languages can I use?
The interface speaks English, Spanish, Italian, French, German and Portuguese, and follows your system language by default. Dictation language is chosen separately; on Windows it can also detect the language automatically.
How can I help?
Try the app, report a reproducible issue, contribute code, or share it with someone who would enjoy it. Optional PayPal support will be available once the creator’s link is configured.