How to run Claude Code, or any AI agent, on a cron schedule

An agent only runs when something calls it. To make one run on a schedule you need a clock outside the agent: a local crontab, a CI scheduler like GitHub Actions, a cloud cron, or a hosted service that calls your endpoint. The choice comes down to one question: what happens when your laptop is closed?

Written by the FacturIQ team. Last updated 2026-09-07. Every claim about FacturIQ here is checkable against the public route table and the public witness log.

The four ways, compared

The simple case: a local crontab

If the work is yours alone and the machine is always on, this is genuinely enough. Claude Code runs non-interactively with -p, so a cron line is all it takes:

# every weekday at 09:00, run one prompt and log the output
0 9 * * 1-5 cd ~/project && /usr/local/bin/claude -p "Review yesterday's commits and write notes to notes.md" >> ~/agent.log 2>&1

Two things to know before you rely on it. Cron runs with a minimal environment, so absolute paths and an explicit PATH matter. And on macOS the machine sleeping means the job does not run; launchd with StartCalendarInterval will catch up after a wake, plain cron will not.

The problem with all of them: silence

Every scheduler above is good at starting a job. None of them, by default, is good at telling you a job did not start. A crontab that stops firing looks exactly like a crontab with nothing to do. The agent does not report the runs it never had, because it was not running.

This is the failure mode that actually costs people work: not a crash, which is loud, but an absence, which is quiet. If a scheduled agent matters, the scheduler needs to record every window, including the empty ones.

A gap in a log has exactly one honest meaning: the scheduler did not run. That is only useful if every attempt writes a line, successes and failures alike.

A wake with a receipt

FacturIQ schedules a signed HTTPS call to an endpoint the agent controls and writes a row for every attempt. Create one with a five-field UTC cron expression:

curl -s -X POST https://api.facturiq.com/v1/wakes \
  -H "authorization: Bearer $FACTURIQ_KEY" -H 'content-type: application/json' \
  -d '{"cron":"0 */6 * * *","webhook_url":"https://your-endpoint.example/wake"}'

Before anything is delivered, the endpoint has to prove it belongs to the agent: FacturIQ posts a challenge and the endpoint answers with a signature made by the agent's own key. After that, each delivery carries an HMAC signature and an Ed25519 signature from the registry, so the receiver can tell a real wake from anything else that finds the URL.

Failures retry after 1, 5 and 25 minutes. Ten consecutive failures disable the schedule and email the operator, rather than retrying into the void. Every attempt, retry, skip and miss is readable:

curl -s https://api.facturiq.com/v1/wakes/12/runs -H "authorization: Bearer $FACTURIQ_KEY"
# {"runs":[{"outcome":"delivered","http_status":200,"duration_ms":142,...},
#          {"outcome":"failed:timeout","attempt":1,...},
#          {"outcome":"missed","scheduled_for":...}]}

If you would rather not run the agent yourself

On the Pro plan FacturIQ can execute the agent instead of calling you: it holds your own Anthropic API key encrypted, runs a short tool loop on the schedule, and records tokens, estimated cost and a summary per run. The model bill goes to your key, not through us. The tool surface is deliberately small, and network access is limited to origins you list, because a scheduled agent reading the open web is a prompt-injection surface.

Which to choose

Use a local crontab when the machine is always on and nobody but you depends on the result. Use GitHub Actions when the work is already in a repository and its secrets belong there. Use a hosted wake when the agent must return whether or not your laptop is open, and when a missed run is something you need to find out about rather than discover weeks later.

Questions people ask