Agent Notch — for Omarchy and macOS

A dynamic-island style notch that hangs under the Omarchy bar (or around the MacBook notch) and shows, live, what your Claude Code sessions (and optionally Grok Bots) are doing — animated faces that think, work, wait for you, celebrate when done and shake on errors. You can also ask Claude something straight from the notch.

▶️ Watch the demo

Which version?

There are two front-ends over the same backend:

Run ./setup.sh and pick Omarchy (Linux, Quickshell plugin) or macOS (native SwiftUI app in macos/). It defaults to whatever matches your machine; ./setup.sh omarchy / ./setup.sh mac skips the question. Both share the same backend (bin/myzk-agents), hooks, config.json and i18n.json. The macOS app needs Swift (xcode-select --install) and supports faces, accessories, moods, pointer reactions and the Customize panel; Omarchy-only bits are placement, screen and the qs ... ipc commands.

On Omarchy the client is a normal window (Agent Notch): open it with the ⤢ button in the expanded notch or ipc call myzk.notch client. It has the same chat list, transcript, composer (agent / folder / model), Git inspector that follows the chat's repo, and the Git & Checksum tab. The Git, checksum and clipboard work is done by bin/agent-notch-tools (Python, uses git, wl-paste and, for the file picker, zenity). Ctrl+V in the notch or the client attaches clipboard images; Esc closes the prompt and keeps the draft.

Features

• Collapsed: an orb for your Claude session on the left, a 2×2 cluster of other agents on the right.
• Hover / click: focus card with a step carousel + one chip per agent.
• Alerts: peeks open on its own when an agent is waiting for input, finishes or fails.
• Ask from the notch: click the Claude orb → type a prompt → runs headless claude -p, the answer renders in the notch (reply, or continue it in a terminal). + cycles the working dir over ~/Projects/*.
• Fully configurable: language (en/es), names, colours, timings, monitor, project dirs — one JSON file, live-reloaded.
• Grok Bots (optional): myzk-agents mcp is a stdio MCP server with a report_status tool.
• Zero dependencies beyond Omarchy (Quickshell) or macOS, and python3.

Install

git clone https://github.com/GNSB/omarchy-agent-notch
cd omarchy-agent-notch
./setup.sh            # or: ./setup.sh omarchy | ./setup.sh mac

Omarchy (./install.sh): it copies the plugin to ~/.config/omarchy/plugins/myzk.notch, the backend to ~/.local/bin/myzk-agents, enables the plugin in ~/.config/omarchy/shell.json, merges the hooks into ~/.claude/settings.json (backups saved as *.bak-notch) and restarts the shell. Uninstall with ./uninstall.sh.

macOS (macos/install.sh): builds the app, installs it as ~/Applications/Agent Notch.app (bundle id com.agentnotch.mac, links ~/.local/bin/agent-notch to it), installs the backend, merges the same Claude Code hooks and registers a LaunchAgent so it starts at login and restarts if it dies. Uninstall with macos/uninstall.sh.

Your own name. By default the backend is myzk-agents, the Omarchy plugin myzk.notch and the state ~/.local/state/myzk-agents. Pick another prefix at install time and everything (backend, plugin id, IPC target, state dir, hooks) follows it:

./setup.sh omarchy --name jorge     # → jorge-agents, jorge.notch, ~/.local/state/jorge-agents
NOTCH_NAME=jorge ./install.sh       # same thing

The name is remembered in ~/.config/agent-notch/name; reinstalling with a different one removes the old plugin and hooks and moves your chats over. In the rest of this README, read myzk as your name.

macOS app

Everything the notch does on Omarchy, plus a full window for when the notch is too small:

• Notch ⇄ client. Click the notch (or agent-notch toggle / agent-notch client, or the Dock icon) and it morphs into the client window; the notch hides. The yellow button (or ⌘M) folds the client back into the notch, which peeks open for a moment. ⌘W closes the client and leaves the notch as usual.
• Chat. Every conversation on the left, the full transcript in the middle (pasted images as thumbnails), and a composer with agent, folder and model pickers. Sessions running in a terminal are read-only here: you get Open in Terminal / New chat here instead of a second process on the same session.
• Models. New chats start on Auto (a cheap Haiku call routes the task to Haiku / Sonnet / Opus, see modelRouter). Pick a model on a chat and it sticks to that chat — across restarts — until you change it.
• Git inspector. The right panel follows the open chat: its working dir if that's a repo, otherwise the repo the chat actually worked in (switch between them under Repos in this chat). Branch, ahead/behind, changed files, remote connectivity and recent commits. It refreshes on its own every few seconds while the client is visible (remotes every minute); ⟳ reloads now.
• Git & Checksum tab. Any repo under projectDirs with the full history, and SHA-256 / SHA-512 / SHA-1 / MD5 of any file, with a box to paste the expected hash and verify it.
• Images. ⌘V attaches images (screenshots, copied images or image files) — paste as many as you like. Clicking away keeps the draft (text + images, the orb shows a badge) so you can grab another screenshot and come back; Esc discards it.
• Permissions. install.sh signs the app with your Apple Development / Developer ID identity when you have one, so macOS remembers the folders you allow (Desktop, Documents…) across rebuilds. Without one it falls back to ad-hoc signing and macOS may ask again after each rebuild. To never be asked, add Agent Notch under System Settings → Privacy & Security → Full Disk Access. Set AGENT_NOTCH_SIGN to pick a specific identity.

agent-notch client | toggle | ask | close | last    # talk to the running app

Chat transcripts live in ~/.local/state/myzk-agents/chats/, per-chat model picks in ~/.local/state/myzk-agents/chat-models.json.

Configure

Everything lives in ~/.config/agent-notch/config.json (created from config.example.json on install). It's live-reloaded — save and the notch updates; no shell restart needed. Missing keys fall back to defaults.

Customize panel. Click ⚙ Customize in the expanded notch (or qs -p /usr/share/omarchy/shell ipc call myzk.notch customize): pick the style, accessories and colour for your assistant or for the Grok Bots, with a live preview you can cycle through every mood and poke. Each click is saved to config.json right away.

Moods & reactions. Faces show idle, thinking, working, upload (Claude runs git push, scp, rsync, npm publish… or an MCP upload), restart (/clear, compaction, resume), waiting, done, error and sleep. They also react to you: the eyes follow the cursor, three quick clicks annoy them, six clicks or shaking the cursor over them make them dizzy. Try it with qs -p /usr/share/omarchy/shell ipc call myzk.notch react dizzy (or annoyed). Grok Bots can report upload and restart through report_status too.

Notch inside the bar ("placement": "bar"): the collapsed notch takes the bar's height and covers its center, so nothing hangs over your browser tabs. Move whatever you have in the bar's center section out of the way first, e.g. omarchy bar move omarchy.clock --section left (repeat for each center widget; the list is under bar.layout.center in ~/.config/omarchy/shell.json). Only top bars are supported.

Texts / translations: all strings are in plugin/i18n.json (shared by the UI and the backend). Add a new language block and set language to it, or override single keys with strings. {name} is replaced with assistantName.

Deeper changes: sizes and animations are in plugin/Notch.qml (widths near the top), the faces in plugin/AgentFace.qml. After editing QML run omarchy restart shell.

Environment overrides: AGENT_NOTCH_CONFIG (config path), AGENT_NOTCH_BACKEND (backend path), AGENT_NOTCH_I18N (strings file for the backend).

Model routing

New asks from the notch or the client (with the model left on Auto) go through a small decision tree in myzk-agents, cheapest step first:

1. rules — free: keywords (es/en), length, file paths, numbered steps and images pick simple, medium or complex when the signal is clear (short = trivially short prompt). Add your own words with modelRouter.words: {"complex": ["\\bkubernetes"]}.
2. cache — the same prompt classified in the last cacheTTL seconds reuses that tier.
3. auto — only what the rules weren't sure about is sent to Haiku to classify.
4. fallback — if the classifier fails, the fallback tier is used.

A routed ask that fails climbs one tier (haiku → sonnet → opus, up to maxEscalations) in a fresh session; the badge then reads escalated ↑. Follow-ups keep the chat's model, and a model picked by hand skips the router. Every routed ask is logged to ~/.local/state/myzk-agents/router.jsonl:

myzk-agents route "refactoriza el módulo de pagos"   # dry run of the rules: tier + signals
myzk-agents router-stats                             # stages, tiers, failures and escalations per starting tier

IPC

qs -p /usr/share/omarchy/shell ipc call myzk.notch toggle   # also: grok, ask, greet, close, last, demo true|false
qs -p /usr/share/omarchy/shell ipc call myzk.notch client   # client window; also: chat KEY, tools, agents

Backend CLI

myzk-agents list | watch                      # see the board in a terminal
myzk-agents set AGENT ID STATE --task "…"     # report from any script (thinking|working|waiting|done|error)
myzk-agents rm AGENT ID

Grok Bots

Add a stdio MCP server in the Grok Bot app: command ~/.local/bin/myzk-agents, args mcp. Bots then call report_status(bot, state, task, detail).

Notes

• Asking from the notch runs claude -p headless with permissionMode (default auto) — set it to default or plan if you want it more careful.
• New translations in i18n.json are very welcome as PRs.
• plugin/record-demo.sh OUT.mp4 records the showcase video with fake agents (needs gpu-screen-recorder).

License

MIT