launchd service setup for local LLM servers

Parent: Mac local LLMs: Serving ops and multi-model · Published reference · snapshot 2026-10-05

LaunchDaemons are scanned and loaded at system start; per-user agents are loaded only when that user logs in. A loaded job starts unconditionally only when RunAtLoad or KeepAlive is set.

These notes link each claim to its source. A source may be a research report hosted on this site rather than the primary document. A published reference means the content is available; it does not certify independent review or accuracy.Read the editorial policy and follow the sources before relying on a claim.

Facts

LaunchDaemons are scanned and loaded at system start; per-user agents are loaded only when that user logs in. A loaded job starts unconditionally only when RunAtLoad or KeepAlive is set. [source]
KeepAlive true restarts the job when it exits, waiting ThrottleInterval seconds between restarts. ExitTimeOut defaults to 20 s (SIGTERM, then SIGKILL). [source]
Modern syntax: `sudo launchctl bootstrap system /Library/LaunchDaemons/x.plist`, `bootout`, and `launchctl bootstrap gui/<uid> ~/Library/LaunchAgents/x.plist`; `load`/`unload` still work but are legacy. [source]
A daemon plist without UserName runs as root, so HOME must be set explicitly or models land in root's home. Working daemon examples set UserName/GroupName, or use a dedicated service account (hidden user `_svcuser`, shell /usr/bin/false, HOME /var/svcuser). [source]
Env vars from a shell profile are not seen by a launchd job; they must be in the plist EnvironmentVariables dict (or in a wrapper script). [source]
Brew-managed plist for Ollama on this box (`sh.brew.ollama`) carries EnvironmentVariables (OLLAMA_FLASH_ATTENTION, OLLAMA_KV_CACHE_TYPE), KeepAlive, RunAtLoad, one combined log file, and LimitLoadToSessionType listing Aqua, Background, LoginWindow, StandardIO, System. [source]
Persisting a `launchctl setenv`: a LaunchAgent with ProgramArguments `/bin/launchctl setenv OLLAMA_HOST <value>` and RunAtLoad re-applies it at every login (present on this box as com.mitch.ollama-host-env.plist). It is a per-login workaround, not a boot-time setting. [source]
A second Ollama instance runs from its own LaunchAgent with a different OLLAMA_HOST port (127.0.0.1:11435), OLLAMA_MAX_LOADED_MODELS and OLLAMA_NUM_PARALLEL in the plist (com.mitch.llmsx-ollama.plist). [source]
oMLX as a daemon: ProgramArguments `/bin/bash -c "exec /opt/homebrew/bin/omlx serve --model-dir ... --host 127.0.0.1 --port ... --paged-ssd-cache-dir ... --hot-cache-max-size 8GB --api-key $(cat <secret file>)"`, ProcessType Interactive, KeepAlive, RunAtLoad, log paths under the service account. [source]
oMLX via Homebrew: `brew services start omlx` (auto-restart on crash); service log `$(brew --prefix)/var/log/omlx.log`, app log `~/.omlx/logs/server.log`; `omlx start|stop|restart` wrap brew services. The formula sets OMLX_SUPERVISED=launchd so the dashboard restart button works under brew services. [source]
oMLX evicts idle models after 5 minutes unless a model is pinned in settings. [source]
Ollama: OLLAMA_KEEP_ALIVE in the plist keeps models loaded (24h in a WoL setup; -1 for a dedicated machine); the default unload is 5 minutes. [source]
mlx_lm.server: the example plist launches a wrapper shell script (ProgramArguments) with RunAtLoad, KeepAlive, a PATH that includes pyenv shims and /opt/homebrew/bin, as a LaunchAgent. [source]
LM Studio headless: llmster installs with `curl -fsSL https://lmstudio.ai/install.sh | bash`, starts with `lms daemon up`; the official startup-task doc is for Linux (systemctl). The desktop app has a "run LLM server on login" setting that minimizes it to the tray. JIT model loading and TTL auto-evict apply to both modes. [source]
Ollama issue 2955 (Mar 2024) asked for a pre-login daemon; the thread produced a working LaunchDaemon (Apr 2024) with UserName, GroupName, EnvironmentVariables OLLAMA_HOST, ExitTimeOut 30, KeepAlive, loaded with `sudo launchctl load`. [source]
Ollama issue 3581 (Apr 2024) shows the documented sequence: quit app, `launchctl setenv`, restart app, check ~/.ollama/logs for "Listening on [::]:11434". [source]
A daemon launched from a script under ~/Desktop (TCC-protected folder) logged "failed with error 1" until the user logged in; moving the binary to /Applications fixed it. Commenter attributes it to the Files and Folders consent system. Keep binaries and models outside Desktop/Documents/Downloads for boot-time daemons. [source]
`brew upgrade ollama` creates a new Cellar dir with a fresh plist, dropping hand edits; copy the plist to ~/Library/LaunchAgents instead (and stop the brew service first) or keep a backup of the EnvironmentVariables block. [source]
metal-guard's shell guard covers interactive shells only, never launchd jobs; launchd scripts must call `metal-guard panic-gate` themselves. [source]
lms daemon: starting the daemon binary and then calling `lms` before it is fully up spawns a second daemon; the two race for llmster-pid.lock and the loser exits 0, which looks like a restart loop under KeepAlive. A wedged auth state was fixed by moving aside ~/.lmstudio/.internal, credentials and settings.json. [source]
A kernel panic reboot plus a plain KeepAlive plist can relaunch the same workload that caused the panic; launchd has no dependency ordering, so a sysctl daemon and the server daemon can race at boot (inferred) [asserted]. [source]
Wake-on-LAN wakes a sleeping Mac only over Ethernet; wake takes 4-7 s, then Ollama needs 3-5 s. [source]
Clamshell/closed-lid laptop dropped to sleep after about an hour despite `pmset sleep 0`; a dummy HDMI plug and a keep-awake app fixed it. [source]
Pre-login daemon on Ollama: issue 2955 reporter could not get anything to run pre-login and suspected disk encryption; another commenter's LaunchDaemon "should start on boot". Both stand: boot start works, but with FileVault on nothing runs until the volume is unlocked. [source]
Sleep prevention: one guide uses `sudo pmset -a sleep 0` plus disksleep 0, tcpkeepalive 1, womp 1, powernap 0; another uses `sudo pmset -a disablesleep 1` for lid-closed laptops. The Astropad guide uses the System Settings energy option instead of pmset. [source]
FileVault on headless Macs: Astropad recommends FileVault off plus auto-login for home use; keep it on in shared spaces; alternatives are `fdesetup authrestart` before planned reboots, an encrypted sparsebundle for secrets, or splitting machines. [source]
Daemon vs agent: Ollama community and the rdner guide use a LaunchDaemon with a service account; the jgoodwill, Hannecke and riclib guides use brew services or LaunchAgents (login required). [source]
Whether a LaunchDaemon with UserName can reach the GPU/Metal pre-login on Apple Silicon is not tested in any source read. [source]
Whether `fdesetup authrestart` survives a kernel-panic reboot (it covers only the next planned reboot per the source). [source]
No source gives the ThrottleInterval default; launchd.info documents the key but only shows 30 as an example. The known default of 10 s is not confirmed here. [source]
A LaunchDaemon from /Library/LaunchDaemons is loaded by the root launchd at system start; LaunchAgents load when the user logs in. [source]
launchd starts a loaded job unconditionally only when RunAtLoad or KeepAlive is set. [source]
KeepAlive true makes launchd wait ThrottleInterval seconds between restarts. [source]
ExitTimeOut defaults to 20 seconds before SIGKILL. [source]
A job's Disabled key can be overridden in /var/db/com.apple.xpc.launchd/disabled.plist; `launchctl print-disabled gui/<uid>` shows it, and bootstrap fails with "Bootstrap failed: 5: Input/output error" if the job is disabled or already loaded. [source]
Use `launchctl bootstrap system <plist>` and `launchctl bootout system <plist>` for daemons, and the gui/<uid> domain for agents. [source]
`launchctl kickstart` restarts a service and `launchctl debug` applies one-shot debug properties that clear after one run. [source]
Setting OLLAMA_HOST in .zshrc does not reach a launchd service; it must go in the plist EnvironmentVariables. [source]
In issue 2955 a commenter could not get anything to execute pre-login and attributed it to disk encryption (decryption on login). [source]
A third-party guide runs Ollama from /Library/LaunchDaemons/com.ollama.plist (no UserName), with HOME, PATH, LANG, OLLAMA_HOST in EnvironmentVariables, logs in /var/log/ollama, owner root:wheel, mode 644, tested on macOS 15.2. [source]
A daemon plist with no UserName runs as root, so ~/.ollama resolves to root's home unless HOME is set (inferred from the guide's explicit HOME key). [source]
A LaunchDaemon with UserName whose binary sat in ~/Desktop failed at boot with open() error 1 until login; installing to /Applications fixed it. [source]
A commenter attributed that failure to the Files and Folders (TCC) protection of Desktop, Documents and Downloads. [source]
brew services keeps Ollama's plist in the Cellar; `brew upgrade ollama` replaces it and drops edits, so back up the EnvironmentVariables block. [source]
Copying homebrew.mxcl.ollama.plist to ~/Library/LaunchAgents survives `brew upgrade`, after stopping the brew service. [source]
`launchctl setenv` values do not survive reboot; plist EnvironmentVariables do. [source]
On this box a LaunchAgent (com.mitch.ollama-host-env.plist) runs `/bin/launchctl setenv OLLAMA_HOST http://127.0.0.1:11434` with RunAtLoad to re-apply the variable each login. [src: local ~/Library/LaunchAgents/com.mitch.ollama-host-env.plist] [source]
On this box the brew Ollama plist `sh.brew.ollama` sets LimitLoadToSessionType to Aqua, Background, LoginWindow, StandardIO, System, and OLLAMA_FLASH_ATTENTION=1 and OLLAMA_KV_CACHE_TYPE=q8_0. [src: local /opt/homebrew/Cellar/ollama/*/sh.brew.ollama.plist] [source]
On this box a second Ollama LaunchAgent serves on 127.0.0.1:11435 with OLLAMA_MAX_LOADED_MODELS=1 and OLLAMA_NUM_PARALLEL=2. [src: local ~/Library/LaunchAgents/com.mitch.llmsx-ollama.plist] [source]
On this box an unrelated server runs as a LaunchDaemon with UserName mitch, GroupName staff, explicit PATH and HOME, RunAtLoad and KeepAlive, logs in the user's home. [src: local /Library/LaunchDaemons/com.openviking.server.plist] [source]
On this box (macOS 27.2) `pmset -g` shows sleep 0 held by sharingd, Claude, powerd, caffeinate and Amphetamine, womp 1, powernap 1, disksleep 10; `sysctl iogpu.wired_limit_mb` returns 0 and FileVault is On. [src: local pmset/sysctl/fdesetup] [source]
A self-hosted oMLX stack runs oMLX, Docling and Open Web UI as LaunchDaemons under a hidden service account `_svcuser` (UID 451, shell /usr/bin/false, HOME /var/svcuser, created with dscl), so services start before login. [source]
That guide's service account needs a password set ("required for launchd") and all installs run through `sudo -u _svcuser`. [source]
The oMLX daemon plist uses `/bin/bash -c "exec omlx serve ... --api-key $(cat /var/svcuser/.config/secrets/omlx)"`, ProcessType Interactive, KeepAlive, RunAtLoad, and plists are mode 644 in /Library/LaunchDaemons, loaded with `sudo launchctl load`. [source]
oMLX evicts models after 5 minutes idle unless the model is pinned in Model Settings. [source]
The oMLX Homebrew formula's brew services job writes both logs to var/log/omlx.log, and since commit 6b2b490 sets OMLX_SUPERVISED=launchd so the admin restart endpoint works under launchd instead of returning 503. [source]
`omlx start|stop|restart` delegate to brew services on Homebrew installs; defaults are ~/.omlx/models and port 8000, and `omlx serve --model-dir X` once persists settings to ~/.omlx/settings.json. [source]
OLLAMA_KEEP_ALIVE=-1 keeps a model loaded indefinitely on a dedicated machine; 24h is used in a wake-on-LAN setup because the default 5-minute unload makes each wake pay 30-60 s model reload. [source]
Wake-on-LAN works only over Ethernet on Macs; measured wake on an M2 Ultra Mac Studio was 4-7 s plus 3-5 s for Ollama. [source]
An MLX server LaunchAgent plist runs a start script with RunAtLoad, KeepAlive, a PATH with pyenv shims and /opt/homebrew/bin, separate stdout and stderr log paths, loaded with `launchctl load`. [source]
llmster (LM Studio headless daemon) installs via `curl -fsSL https://lmstudio.ai/install.sh | bash` and starts with `lms daemon up`; the official startup-task doc covers Linux systemctl only. [source]
LM Studio's desktop-app headless mode is a login setting that keeps the server running from the tray; `lms server start` restores the last server state programmatically; JIT-loaded models auto-unload after a TTL. [source]
With llmster, calling `lms` before the daemon has finished booting spawns a second daemon, and the two race for llmster-pid.lock; the loser exits 0 ("Exiting due to PID lock loss"), which looks like a restart loop in a supervisor. [source]
llmster auth state wedged after daemon churn; moving aside ~/.lmstudio/.internal, credentials and settings.json fixed it while keeping models and extensions. [source]
metal-guard's shell guard covers interactive terminals only and never launchd jobs or scripts; `metal-guard panic-gate` is the command meant for launchd scripts. [source]
metal-guard's L10 and L12 layers handle recovery after a panic and reboot, and a persistent MLX worker subprocess respawns on crash. [source]
With FileVault on, a headless Mac needs a password at the pre-boot screen before the network is up, so after a power flicker, update or kernel panic it is unreachable until someone types it locally. [source]
Astropad's guidance for a home headless Mac mini is FileVault off and Automatic login on, so it recovers after a power outage. [source]
`sudo fdesetup authrestart` before a planned reboot unlocks the disk once on next boot; the guide adds that on macOS 26 with Ethernet one can SSH in at the pre-boot stage to unlock. [source]
The login keychain stays encrypted with the user password even with FileVault off. [source]
Automatic login does not work with FileVault enabled. [source]
`sudo pmset -a autorestart 1` restarts the Mac after a power failure. [source]
A headless sleep-prevention set is `pmset -a sleep 0`, `disksleep 0`, `tcpkeepalive 1`, `womp 1`, `powernap 0`; `pmset -g` may show "sleep 0" with a note naming what holds sleep off rather than a SleepDisabled line. [source]
A MacBook Pro in closed-lid mode still slept after about an hour with `sleep 0`; a dummy HDMI display emulator and a keep-awake app fixed it. [source]
`sudo pmset -a disablesleep 1` prevents sleep with the lid closed; undo with `disablesleep 0`; closed-lid running needs airflow. [source]
The Astropad guide sets "Prevent automatic sleeping when the display is off" in Energy Saver rather than pmset. [source]
A launchd watchdog pattern for 24/7 agents: a script run every 2 minutes and at load, PID-file lock, a kill-switch file, memory-pressure check, heartbeat-staleness zombie detection. [source]
launchd has no ordering between independent jobs, so a boot-time sysctl daemon and a server daemon can start in either order; a server wrapper should run `sysctl` itself or wait on it. [source]
A server job under launchd after a panic reboot should call a panic gate before starting the model, otherwise KeepAlive relaunches the same workload (extends the 14-minute respawn claim). [source]

Corrections and disagreements

CONTRADICTS: ollama-on-macos.md line 28 and 77. A working Ollama LaunchDaemon exists in issue 2955 (UserName, GroupName, OLLAMA_HOST in EnvironmentVariables, ExitTimeOut 30, KeepAlive, `/opt/homebrew/bin/ollama serve`, installed with `sudo cp ollama.plist /Library/LaunchDaemons/`) and is reported to start on boot. [source]

Children

No children recorded.

← the whole tree · 3D view· how to read this page