The agent is still working.
Don’t let the Mac quit.

LidGuard watches Claude Code, Codex and the rest, and holds your Mac awake only while there is real work happening. When macOS won't honour that, it says so instead of pretending.

Read what it actually does $1 once, at launch. macOS 14 or later.

Not released yet. The build is signed and going through Apple’s notarization. Homebrew install and the $1 licence arrive with it.

LidGuard

Awake — lid-proof · 0h 13m

Battery

58% · 2h 55m leftstops below 15%

Agents1 working

Claude Codelifecycle hooks 1 session

OpenAI Codex CLIprocess detection idle

OpenCodeprocess detection idle

Pause 30 min1 hour

Settings…⌘,

Quit LidGuard⌘Q

What macOS actually allows

Every keep-awake app uses the same power assertions. None of them stop the sleep that closing the lid triggers on battery — that needs root. Here is the whole picture, measured on macOS 26.

Hold	Mechanism	Admin password	Stops idle sleep	Stops lid-close sleep on battery
Standard	`PreventUserIdleSystemSleep` `PreventSystemSleep`	Never	Yes	No
Deep hold	`pmset -a disablesleep 1`	Once, at install	Yes	Yes

Deep hold is opt-in and installs a small helper. Until you install it, the menu reads lid open rather than lid-proof. The helper carries a dead-man switch: LidGuard refreshes a timestamp every 30 seconds, and if that goes stale for two minutes the helper re-enables sleep on its own. A crash can't leave your Mac unable to sleep.

Running is not the same as working

An agent sitting at a prompt is not an agent mid-task. LidGuard uses two mechanisms and always shows you which one is in play for each agent.

Lifecycle hooks

A loopback-only endpoint that agents post to. A prompt opens a turn, Stop closes it, a permission prompt counts as waiting on you. This is a true working/idle edge, per session, so two windows can be busy and quiet independently.

Installed into ~/.claude/settings.json by merging — your existing hooks are preserved, and a backup is written first.

Process detection

For agents with no hooks. LidGuard matches path components of each process's real executable, which is the only thing that works: Claude Code's binary lives at ~/.local/share/claude/versions/2.1.263, so its filename and its process name are both a version number.

With no hook signal a live agent counts as working, because sleeping mid-task is the failure worth avoiding.

Claude Code — hooks
Codex CLI — process
OpenCode — process
Gemini CLI — process
Copilot CLI — process
Aider — process
Cursor — process
Cline — process

Any other agent can be wired to the same endpoint with a one-line curl in its own hook config.

The rest of it

Battery guardrails — a cut-off percentage, an only-while-plugged-in mode, and Low Power Mode respect.
Display control — keep the screen on, turn it off while agents run, or turn it off a set time after they finish.
Notifications and a chime when a hold engages, when agents finish, or when a guardrail stops it.
Pause for 30 minutes or an hour without turning anything off.
Toggle everything with ⌥⌘L, registered without needing Accessibility access.

The agent is still working.Don’t let the Mac quit.