Know the moment an agent turns to you.

The Eye is a small menu-bar app on macOS and system-tray app on Windows that keeps watch over the AI coding agents you run. It stays dark and quiet while they work, opens red with a count the instant one stops to ask you something, and rests again when all the work is done, so you never have to keep flipping back to a terminal to check.

Claude Code, OpenAI Codex CLI, Cursor (CLI and in-editor), GitHub Copilot (CLI and in VS Code), and Google Antigravity (CLI and in-editor). One install brings them all under the same gaze, in the terminal or your editor, several at once.

When you run several agents in parallel, you stop babysitting their windows. The Eye gives you one place to glance: who's calling for aid, who's still working, and who's gone idle. Click a waiting session and the exact terminal or editor window comes forward. No hunting through tabs. You can also be told by a sound, a notification, a Telegram message to your phone, or by bringing the window forward on its own, and each of those is opt-in.

Yes, through Telegram. The Eye can send you a message the moment an agent starts waiting, so you're told even when you're away from your desk. You connect your own Telegram bot once from Settings → Connectors, with step-by-step guidance built in; it takes about a minute. Your bot token is kept in the macOS Keychain (in the app's local preferences on Windows), the message carries only the agent and project name (never your code), and nothing is sent until you turn it on.

It listens for the lifecycle hooks each agent already emits. A one-time setup from Settings adds a tiny status hook to each agent's own config; as an agent starts, works, finishes, or stops to ask you, that hook writes a small local status file on your machine. The Eye watches those files and reflects the combined state in the menu bar or tray the instant anything changes. On Windows, when Codex does not invoke hooks, The Eye reads the same lifecycle state from Codex's local rollout log instead.

It watches from the side and never runs, drives, or controls your agents. They keep running in their own apps exactly as before.

No, and that is its core design decision. It watches from the side: your agents keep running in their own tools exactly as before, and The Eye never wraps, drives, or restarts them. Existing tools wanted to own the whole workflow; The Eye lets you stay in your terminal and editor.

Orchestrators run your agents for you: they wrap the agent process, own the workflow, and give you a dashboard to keep open. The Eye does the opposite. It never launches or controls anything; it only reads the lifecycle signals your agents already emit, and reports them in one light in the menu bar or tray. You keep your own terminal, your own editor, and your own way of working. The tradeoff is deliberate: The Eye cannot start or retry an agent for you, and in exchange it cannot get in the way of one.

Your code, prompts, and output never leave your machine. The status files are plain local files that carry only what's needed to show state: a session name, its tool, and whether it's working or waiting. On Windows, when Codex does not invoke hooks, The Eye reads lifecycle records and the final assistant message from Codex's local rollout log only to determine whether that session is working or waiting; that content stays on your machine and is not copied into analytics. There's no account and no server of ours to send your work to.

The macOS app is signed with an Apple Developer ID and notarized by Apple, so Gatekeeper opens it with no warning, and every update verifies its signature before it installs; the Windows beta is not code-signed yet, so SmartScreen warns once on first run. On macOS and Windows it sends nothing on its own except optional, anonymous usage counts (app launches and periodic heartbeats), which you can turn off under Settings or Preferences → Privacy. Those counts include the app version, OS, and CPU architecture, never agent names, paths, prompts, output, or code. If you switch on Telegram notifications, it also sends a short alert, the agent and project name only, to your own bot; that stays off until you set it up.

Yes, as a beta. Version 1.2.1 is a system-tray app for Windows 10 and 11: the same watch light, the same agents, the same local status files, the same one-button hook setup under Preferences → Hooks, a draggable always-on-top status window, and in-app feedback. It is not code-signed yet, so SmartScreen warns once on first run — choose More info, then Run anyway; it installs for your user only, so there is no admin prompt. It uses the same optional anonymous usage counts as macOS, with an opt-out under Preferences → Privacy.

It's free. On macOS it needs 14 Sonoma or later and is a universal build, native on both Apple Silicon and Intel Macs; the Windows beta needs 64-bit Windows 10 or 11. Download the app, drag it to Applications (or run the Windows installer), and open Settings → Tools, where one button sets up every agent it finds. It takes about a minute, and there's nothing to run in a terminal.

On macOS, signed and notarized updates arrive quietly in the background, and you can also check on demand from About → Check for Updates; those updates verify both a cryptographic key and the app's Developer ID signature. From 1.2.0, the Windows beta checks daily too: it tells you a new version is out — a notification and an item in the tray menu — and installs only when you click Install under Preferences → General → Updates, verifying the download against a cryptographic key first. Copies older than 1.2.0 need the newer installer run over them once.

In Settings → Tools, Remove all takes back every hook The Eye added to your agents' configs, leaving the rest of your settings untouched. After that, quit The Eye and move it to the Trash on macOS, or uninstall it from Windows Settings → Apps.