By name: scheduling description: | Set up a recurring or scheduled task — "every morning", "on a cron", "remind me daily", "run this report weekly", "keep checking X". Decides between a Claude routine (each run needs judgment) and a macOS LaunchAgent (deterministic code), and sets up whichever fits. version: 1.0.0
Scheduling Recurring Tasks
When the operator wants something to run on a schedule, pick the mechanism with one question, then set it up for them.
Pick: does each run need Claude to think?
Yes — it needs judgment (summarizing, drafting, deciding, reading natural
language, reacting to fuzzy or changing conditions):
→ Claude routine. A saved Claude Code config that runs autonomously on
Anthropic's cloud on a schedule. Create it with /schedule (e.g. /schedule daily PR review at 9am). It runs even when the operator's laptop is closed, and
each run shows up as a session at claude.ai/code.
No — it's the same deterministic steps every time (a sync, a backup, a
git pull, generating a report from a script, pinging an API):
→ macOS LaunchAgent. A local job that runs as the operator while they're
logged in.
Two things that override the simple pick
- Cost. A Claude routine draws down the operator's Claude subscription usage exactly like a normal session does — it is not free cloud compute — and there's a daily cap on routine runs per account. A LaunchAgent running a plain script (no Claude call) costs nothing. So for fixed, deterministic work, especially anything frequent, use a LaunchAgent.
- Cadence & locality. Routines can't run more often than hourly and run in the cloud (no access to the operator's local files — they clone a GitHub repo instead). LaunchAgents go down to every minute and can touch anything on the machine. So: sub-hourly, or needs local files/apps → LaunchAgent.
Quick examples
| Request | Use | Why |
|---|---|---|
| "Each morning, draft my standup from my calendar + Slack" | Claude routine | needs judgment, low frequency, uses connectors |
| "Every hour, pull the latest Traba skills" | LaunchAgent | fixed code, free, local |
| "Every 5 min, sync this file / poll this API into a CSV" | LaunchAgent | fixed and sub-hourly |
| "Weekly, summarize merged PRs and open doc-update PRs" | Claude routine | judgment, repo-centric, weekly |
/loopand in-session cron run only while a session is open — live, ephemeral repetition, not durable scheduling. For "set it and forget it," use a routine or a LaunchAgent.
Set up a Claude routine
Requires a Pro / Max / Team / Enterprise plan with Claude Code on the web enabled.
Run /schedule and describe the task and cadence — e.g. /schedule every weekday at 8am, summarize yesterday's merged PRs. Claude collects the prompt, repo(s),
connectors, and schedule, then saves it. Confirm the parsed run time with the
operator and give them the claude.ai/code URL where runs appear. Keep the cadence
low — every run spends subscription usage.
Set up a LaunchAgent
- Write the script the job runs (bash/python/node) with absolute paths —
launchd runs with a minimal environment, so don't rely on the operator's shell
PATH or aliases. Make it executable (
chmod +x). - Write the plist to
~/Library/LaunchAgents/work.traba.TASK_NAME.plist. Start fromlaunchagent.template.plistin this skill. Key choices:StartCalendarIntervalfor specific times (like cron);StartIntervalfor "every N seconds".- Set
StandardOutPath/StandardErrorPathso there's a log to debug from.
- Load it:
launchctl bootstrap gui/$(id -u) ~/Library/LaunchAgents/work.traba.TASK_NAME.plist # older macOS: launchctl load ~/Library/LaunchAgents/work.traba.TASK_NAME.plist - Tell the operator what you set up, when it runs, where the log is, and how to stop it.
Change or remove it:
launchctl bootout gui/$(id -u)/work.traba.TASK_NAME # stop it
# edit the plist, then bootstrap again to reload
rm ~/Library/LaunchAgents/work.traba.TASK_NAME.plist # delete entirely
Gotchas:
- Run the script directly first to confirm it works — a job that errors on schedule is invisible unless you set a log path.
- Absolute paths everywhere, in both the plist
ProgramArgumentsand the script. - If the Mac is asleep at the scheduled time, launchd runs the job when it next wakes; it doesn't fire once per missed slot.
Always tell the operator what you chose
State which mechanism you used and why — e.g. "this runs the same steps every time and touches local files, so I set it up as a local LaunchAgent; it costs nothing and runs every hour" — so they understand it and can manage it later.