wakivo

Keep your machine awake while your work runs. Then let it sleep again.

Built for the case where you start something long (an agentic coding session, a training run, a big build), walk away from your desk, and come back to find the machine went to sleep half way through.

wakivo -- uv run train.py     # stay awake until the command exits
wakivo --pid 41823            # stay awake until that process exits
wakivo 2h                     # stay awake for two hours
wakivo                        # stay awake until Ctrl-C

wakivo exits with the wrapped command's own exit code, so it drops into scripts and CI without changing their behaviour.

Install

uv tool install wakivo            # the CLI
uv tool install "wakivo[gui]"     # and the menu bar app

pipx install works the same way. Requires Python 3.13+.

Menu bar app

wakivo-gui puts a cup in the menu bar (macOS) or notification area (Windows), for when you would rather not open a terminal.

● Awake — 1h 30m left
──────────────────────
Keep awake for      ▸   15m · 30m · 1h · 2h · 4h
Keep awake until I quit
Stop
──────────────────────
☑ Also keep the display on
Quit

Two modes, on purpose. Binding a hold to a particular process means knowing which process, and anyone who does is already served by wakivo --pid — so the GUI has no picker to misunderstand. The icon is filled while a hold is active and outlined when it is not, so "is it on?" needs no click.

Launching it without a terminal

wakivo-gui --install       # then launch it like any other app
wakivo-gui --uninstall

On macOS this writes ~/Applications/Wakivo.app; on Windows, a Start menu shortcut. Run it once and you never need a shell again.

No code signing is involved, because the launcher is built on your machine rather than downloaded. Gatekeeper and SmartScreen act on the quarantine attribute, which is applied by whatever fetched a file from elsewhere — a bundle your own computer just wrote never carries one. Nothing here embeds a Python runtime either; the launcher simply starts the interpreter you already have.

macOS and Windows only. Linux keeps the CLI: GNOME removed system tray support years ago, so a tray icon needs a shell extension the user has to install first. wakivo-gui says so and exits rather than half-working.

Options

Durations accept 90s, 20m, 2h, 1h30m, or a bare number of seconds.

Scope

wakivo prevents idle sleep — the machine suspending because you stopped touching it. That is the failure that interrupts a long task. Everything else is deliberately out of scope.

A blanked, locked screen with your job still running is the intended outcome, not a failure — it is why screen-off is the default rather than -d. Neither blanking nor locking stops a running process.

Two of those rows are worth explaining, because they surprise people:

Closing the lid still sleeps the machine. The lid switch is a hardware event, separate from idle sleep, and no wakelock on any platform survives it — not caffeinate, not SetThreadExecutionState, not wakivo. Holding a laptop awake through a closed lid requires root and a persistent change to system power settings, so a crash could leave your machine permanently unable to sleep. That trade is deliberately not made here. It may return as an explicit opt-in flag.

A critical battery still sleeps the machine. UPower on Linux, and the equivalent elsewhere, suspends or hibernates directly at critical charge and ignores every inhibitor by design. You cannot inhibit your way out of a dying battery, and shouldn't be able to.

Design

Two independent axes:

• backends/ — how the hold is taken, per platform. macOS uses IOKit power assertions via ctypes; Windows uses SetThreadExecutionState on a dedicated parked thread, because the flag is per-thread and evaporates when the setting thread exits; Linux takes a systemd-logind inhibitor, plus GNOME's own session inhibitor when running inside a GNOME session.
• triggers/ — when the hold is released: a command exiting, a pid dying, a timer, or an interrupt.
• session.py — a hold that can be started and stopped rather than waited out. The CLI runs to completion; the menu bar app needs to acquire and return, then release later. Both drive the same backends and triggers.

Any trigger composes with any backend, so a new release condition is a new file rather than a change to the wakelock code.

Every backend must be crash-safe: the OS drops the hold when the process dies, including on SIGKILL. wakivo never mutates persistent system settings, so there is nothing to clean up and nothing to restore.

Why Linux drives a subprocess

macOS and Windows call the platform API directly rather than shelling out to caffeinate — a child process is one more thing that can outlive us still holding the hold.

logind is genuinely different. Its inhibitor is handed out as a file descriptor over D-Bus, so taking it in-process means implementing D-Bus fd passing. systemd-inhibit is the canonical client for that, so the Linux backend runs it as a child and uses PR_SET_PDEATHSIG to keep the guarantee: the kernel kills the helper the moment wakivo dies, so the descriptor is dropped however we exit. Taking the fd directly would remove the child entirely and is worth doing later.

Two holds on Linux, not one

A logind inhibitor alone does not stop GNOME. Measured on Debian/GNOME with a 60-second idle timeout, journal markers either side of the hold:

18:51:41  CONTROL START
18:52:41  will suspend now      <- 60s, control valid
18:54:26  HOLD START
18:56:18  will suspend now      <- 112s, suspended mid-hold
18:59:31  HOLD END

wakivo was holding sleep:idle in block mode throughout — strong enough that systemctl suspend was refused outright — and gsd-power suspended the machine anyway. GNOME runs its own idle policy against its own session inhibitors, which live on the session bus and are entirely separate from logind's. Adding gnome-session-inhibit fixed it: the same run now reaches HOLD END with no suspend in between.

So inside a GNOME session wakivo takes both: the logind inhibitor, which governs headless and non-GNOME systems, and gnome-session-inhibit, which is the one GNOME actually consults. Neither subsumes the other — a Debian server has no gnome-session at all.

The logind hold tries --what=idle:sleep first and falls back to idle. Blocking sleep also stops Suspend() calls, covering desktops whose power daemon hasn't been measured here — but it needs the org.freedesktop.login1.inhibit-block-sleep polkit action, which is denied without an active seat session, so headless servers and CI runners are refused outright. Blocking idle is granted broadly and covers logind's own IdleAction, which is exactly what governs those machines.

Where the stronger hold is permitted, a deliberate systemctl suspend is refused while wakivo runs. That differs from macOS, where PreventUserIdleSystemSleep leaves intentional sleep alone.

The GNOME hold inhibits suspend only, deliberately not idle — GNOME's idle inhibitor also suppresses screen blanking and locking, and screen-off is wakivo's default.

Status

macOS, Windows and Linux all pass in CI, each asserting against its own power tooling — pmset -g assertions, powercfg /requests, systemd-inhibit --list — that the hold is visible to the OS and is reclaimed when the process is killed outright.

Beyond CI, all three have been verified on real hardware the only way that counts: shorten the idle timeout to one minute, confirm with a control run that the machine does sleep, then hold for four minutes and confirm it doesn't.

Two honest limits on what CI itself proves:

• No CI runner ever idles into sleep, so CI shows the platform accepts the hold and reports it, not that the machine stays awake. That part is only verifiable by hand.
• Linux runs on ubuntu-latest, a headless systemd VM with no seat session. It covers the logind path and the fallback to an idle-only hold, but never reaches the GNOME branch — which is exactly where the bug above lived. Desktop coverage is manual.

--keep-display is unimplemented on Linux; it needs the DE-specific screensaver interfaces, and is refused rather than silently ignored.

Development

uv sync
uv run pytest

The backend tests take a real wakelock and assert the OS can see it, so they catch a broken ctypes signature before a user does. On macOS one of them SIGKILLs a child holding an assertion and checks the kernel reclaimed it.

Licence

Apache-2.0 — see LICENSE and NOTICE.

Chosen over MIT for two things it adds. It grants patent rights explicitly, so anyone building on this is not relying on an implied licence. And anything redistributing a substantial portion has to carry the contents of NOTICE forward, which is a real attribution requirement rather than a copyright line that can sit unread in a bundled licence file.

It stays permissive: you may use, modify, redistribute and sell this, and you are not obliged to publish your changes.