Day 3: Set It Free

Until now, your bot only talks when you talk first. Today you give it a schedule: automations that run at fixed times (a morning briefing, an hourly check), and the heartbeat, a periodic check-in that stays quiet unless something needs your attention.

Time: ~45 minutes

Automations and the Heartbeat

OpenClaw has two kinds of scheduled work, and both are run by the Gateway's automations scheduler (automation overview):

	Automations	Heartbeat
What it is	A job you create: a schedule plus a prompt (or a command)	One system-owned monitor job per agent
When it runs	Exactly when you say: 0 8 * * *, every 2h, a one-shot --at 20m	Every 30m by default (1h with Anthropic OAuth/token auth)
Where it runs	A fresh isolated session when created from the CLI; the creating conversation when created in chat	The agent's main session by default
What it does	Whatever the prompt says, then delivers the result	Checks its scratch notes and replies NO_REPLY if nothing needs attention
Use it for	Briefings, reports, recurring checks	"Is anything urgent?" sweeps, follow-ups on background work

Rule of thumb: if the task has a time or interval in its name, it's an automation.

Coming from an older guide?

Older OpenClaw versions read heartbeat tasks from a HEARTBEAT.md file in the workspace. That file was retired in 2026.8.1: OpenClaw no longer creates or reads it. If you still have one, stop the Gateway (openclaw gateway stop) and run openclaw doctor --fix. Doctor imports its instructions into the heartbeat's monitor scratch, turns valid tasks: entries into automations, and archives the original (retired HEARTBEAT.md).

Before You Start: Where Messages Go

Scheduled work needs somewhere to send its results.

• Heartbeat alerts go to the owner's direct message, taken from commands.ownerAllowFrom (or a concrete channel allowFrom). Without an owner, heartbeat polls are skipped with reason=no-route. If you approved your own first Telegram DM with openclaw pairing approve, the owner is already set. Otherwise set it yourself, using your numeric Telegram user ID:

  openclaw config set commands.ownerAllowFrom '["telegram:123456789"]'

• Automations you create from a chat reply to that chat. From the CLI, a job defaults to --channel last, which resolves from session history (or your only configured channel) and fails closed if it can't. Pin the destination with --channel telegram --to "<chat id>" (for a DM with your bot, the chat ID is your user ID); openclaw cron show <jobId> previews the resolved route.

• Times use your timezone. Set agents.defaults.userTimezone (see Day 1); automation cron expressions use the Gateway host's timezone unless you pass --tz.

Your First Automation

The easiest way is to ask in chat:

Every weekday at 8am, send me the weather for Berlin and the top 5 Hacker News stories about AI.

The agent creates the automation with its automations tool and replies with the job details. The same job from the CLI:

openclaw cron add "0 8 * * 1-5" \
  "Send me the weather for Berlin and the top 5 Hacker News stories about AI, one line each." \
  --name "Morning brief" \
  --tz Europe/Berlin \
  --tools web_search,web_fetch,read \
  --channel telegram --to "123456789"

• The first argument is a standard 5-field cron expression; the second is the prompt.
• --tools caps what the job may use. A briefing that only reads the web has no reason to run shell commands. Without --tools, each run inherits the agent's full tool policy.
• openclaw automations is the same command as openclaw cron; use whichever reads better.

Email and calendar items need a skill that can reach them, such as the bundled gog (Google Workspace) skill. You'll set that up on Day 4.

tip

Test it now instead of waiting until morning. Find the job ID, then run it once and wait for the result:

openclaw cron list
openclaw cron run <jobId> --wait

Practical Automation Ideas

Each idea below is one command. Adjust the schedule, prompt and delivery target, or just describe the job to your bot in chat.

Morning Briefing

openclaw cron add "0 8 * * *" \
  "Weather forecast for Berlin and the top 5 Hacker News stories about AI." \
  --name "Morning brief" --tools web_search,web_fetch,read \
  --channel telegram --to "123456789"

GitHub Repo Monitoring

openclaw cron add "every 2h" \
  "Check github.com/myorg/myrepo for issues and PRs opened in the last 2 hours. Categorize each as bug, feature or question. Say if CI is failing on main. If there is nothing new, reply NO_REPLY." \
  --name "Repo watch" --tools web_fetch \
  --channel telegram --to "123456789"

Issue and PR text is written by strangers. Treat it as untrusted input: keep this job's tools to read-only web access, and don't give it a token that can merge or push. A NO_REPLY answer is not delivered, so quiet runs don't message you.

Server Health Check

For checks with a clear pass/fail rule, a command job is safer and cheaper than a model run. It executes a fixed shell command on the Gateway host, with no model in the loop, and delivers its output. Output that is exactly NO_REPLY is suppressed. The command runs inside the Gateway process, not through the agent's exec tool, so exec approvals do not apply: only an operator (operator.admin) can create one, and the model can't change what it runs:

openclaw cron add "0 * * * *" \
  --name "Disk check" \
  --command 'p=$(df --output=pcent / | tail -1 | tr -dc 0-9); if [ "$p" -gt 85 ]; then echo "Root disk at ${p}%"; else echo NO_REPLY; fi' \
  --announce --channel telegram --to "123456789"

df --output is GNU coreutils (Linux). On macOS, adapt the command. To check a remote server, run this script on that server instead of giving your bot an SSH key; the server monitoring recipe shows the guardrails if you do want agent-driven SSH.

Price Monitoring

openclaw cron add "every 2h" \
  "Check the current BTC and ETH prices. Alert me only if BTC moved more than 5% since the previous check; otherwise reply NO_REPLY." \
  --name "Price watch" --tools web_search,web_fetch \
  --channel telegram --to "123456789"

Each run is isolated by default, so "since the previous check" only works if the agent can see earlier results. Add --session session:price-watch to give the job a persistent named session that keeps its history across runs, or have the prompt compare against a fixed threshold instead.

Daily Journal Prompt

openclaw cron add "0 21 * * *" \
  "Ask me: What was the most interesting thing that happened today? And: What's one thing you want to do differently tomorrow?" \
  --name "Journal prompt" --tools "" \
  --channel telegram --to "123456789"

--tools "" disables all tools: this job only needs to send a message. Reply in your normal chat and ask the bot to save your answers to journal/ in the workspace.

Weekly Summary

openclaw cron add "0 18 * * 0" \
  "Read this week's daily notes in memory/ and list the top 3 things I worked on, plus one suggestion for next week." \
  --name "Weekly summary" --tools read,memory_search \
  --channel telegram --to "123456789"

The Heartbeat Monitor

The heartbeat is on by default. Every 30m it runs a short turn in the agent's main session with this built-in prompt: follow the monitor scratch if there is one, keep recurring work in automations, don't dig up old tasks from earlier chats, and reply NO_REPLY if nothing needs attention (heartbeat). On a fresh install it mostly stays silent.

Tune it in agents.defaults.heartbeat:

~/.openclaw/openclaw.json

{
  commands: {
    ownerAllowFrom: ["telegram:123456789"],
  },
  agents: {
    defaults: {
      userTimezone: "Europe/Berlin",
      heartbeat: {
        // How often it runs: a duration string. "0m" turns off the recurring cadence.
        every: "1h",
        // When it may run, in userTimezone. Outside this window, ticks are skipped.
        activeHours: { start: "07:00", end: "22:00" },
        // Skip the workspace bootstrap files on heartbeat runs.
        lightContext: true,
        // Fresh session each run instead of your full chat history.
        isolatedSession: true,
      },
    },
  },
}

Run openclaw config validate after editing. Unknown keys stop the Gateway from starting. The Gateway applies heartbeat changes on config reload, or run openclaw gateway restart.

activeHours says when the heartbeat may run. That is the opposite of the "quiet hours" setting in some older guides. Scheduled heartbeats also need automations enabled; they stop if cron.enabled is false.

Give the heartbeat a checklist

The heartbeat's instructions live in its monitor scratch, a small private note stored in the Gateway's database. Find the heartbeat job (it is listed as Heartbeat (main)) and set the scratch:

openclaw cron list --all
openclaw cron scratch <heartbeatJobId> --set "If a background job I started has finished, tell me the result. Remind me of any reminder I asked for that is now due."

Keep scratch short and about checking, not scheduling. Anything with a time in it ("at 8am", "every Monday") belongs in an automation.

Model Choice and Cost

Don't put tool-using scheduled jobs on your cheapest model. Automations and heartbeats run unattended, and the jobs above read web pages and issues written by other people. Upstream's security guidance is explicit: smaller, cheaper models are more susceptible to prompt injection and tool misuse, so use a current top-tier model for any agent that can run tools or reads untrusted content (prompt injection). A small model is fine for a job with --tools "" that only formats a message.

Control cost with scope instead:

1. Run less often. The heartbeat's every: "1h" halves the runs of the 30-minute default, and activeHours skips the night.
2. Send less context. A heartbeat in the main session carries your bootstrap files and conversation history on every run. lightContext and isolatedSession (above) shrink each run considerably. For automations, --light-context does the same.
3. Use command jobs for anything a script can decide. They make no model call.
4. Narrow tools. --tools keeps each job's tool list, and its tool schemas, small.

See what you are actually spending:

openclaw gateway usage-cost --days 7

In chat, /usage cost shows the same for the current session. For model options and prices, see the Model Selection Guide.

Managing Automations

Automations take a job ID, shown by openclaw cron list:

# List enabled jobs (add --all to include disabled ones and the heartbeat)
openclaw cron list

# Show one job, including its resolved delivery route
openclaw cron show <jobId>

# Pause and resume
openclaw cron disable <jobId>
openclaw cron enable <jobId>

# Delete
openclaw cron rm <jobId>

openclaw cron edit <jobId> changes the prompt, schedule, tools or delivery of an existing job. See the automations CLI for every option.

Debugging

If a scheduled job or the heartbeat isn't doing what you expect:

# Is the scheduler running, and when is the next run?
openclaw cron status

# Run history for one job (status, errors, delivery)
openclaw cron runs <jobId>

# Run a job right now and wait for it
openclaw cron run <jobId> --wait

# What did the last heartbeat do?
openclaw system heartbeat last

# Wake the heartbeat now with a one-off note
openclaw system event --text "Check whether my backup finished" --mode now

# Stream Gateway logs
openclaw logs --follow

Common issues:

• Nothing runs: Check that openclaw gateway status shows the Gateway running, and openclaw cron status shows the scheduler enabled. cron.enabled: false or OPENCLAW_SKIP_CRON=1 turns scheduling off.
• Heartbeat never messages you: That is normal when nothing needs attention (NO_REPLY). If the logs show heartbeat polls skipped with reason=no-route, there is no owner to deliver to: set commands.ownerAllowFrom.
• Automation runs but you get nothing: Check the delivery column in openclaw cron runs <jobId> and the route in openclaw cron show <jobId>. Pin it with openclaw cron edit <jobId> --channel telegram --to "<chat id>".
• Wrong timing: Check --tz on the job and agents.defaults.userTimezone. Heartbeats outside activeHours are skipped.
• High costs: Lower the frequency, add lightContext/isolatedSession, and move script-decidable checks to command jobs. Check openclaw gateway usage-cost.

For the full technical details, see the Heartbeat Guide, Heartbeat Architecture, and upstream's automation troubleshooting.

---

Next: Day 4: Supercharge It — install skills to give your bot new abilities.