Can Hermes check something every few minutes and tell me only when it changes?
Yes. In the conversation where you want the answer, send /loop 5m <what to check> or /heartbeat every 10m <what to check>, and Hermes re-runs that instruction on a timer while the conversation is idle. To hear only about changes, ask for a bare [SILENT] reply when nothing changed. On a messaging channel such as Telegram, Hermes doesn’t deliver that reply. Each check is still a full model turn in the same conversation, though, so a quiet check costs nearly as much as a noisy one. If the check is a page, a price or a status you can read with a script, a scheduled job with a pre-check script is cheaper.
This follows Hermes Agent v0.23.1, commit 3b42157, checked on 18 November 2026. We reviewed official documentation and source. We did not run a watch on a Vavio computer or measure delivery or cost.
How do I start a watch?
Use /loop for most watches. It can stop itself when the thing you’re watching is finished:
/loop 5m Check the tracking page for order 4471. If the status is the same as last time, reply with only [SILENT]. If it changed, tell me the new status.
| Command | What it does |
|---|---|
/loop 5m <prompt> |
Checks every 5 minutes. Intervals such as 30s, 2h and 1h30m work; anything shorter than 30 seconds is raised to 30 |
/loop <prompt> |
Self-paced: starts at 1 minute and doubles the wait while replies stay the same, up to 15 minutes; any change goes back to 1 minute |
/loop 5m <prompt> --times 12 |
Stops after 12 checks |
/loop 5m <prompt> --until "the parcel is delivered" |
Stops when a judge model decides the condition is met, and pauses if it judges it can’t be |
/heartbeat every 10m <prompt> |
A standing instruction with no end. Intervals under 60 seconds are refused, and 1d works |
/loop status, pause, resume, stop |
/heartbeat has the same controls; clear removes either |
Each conversation has one loop and one heartbeat; setting a new one replaces the old. The first /loop check is due straight away (on a messaging channel, within about 15 seconds); the first heartbeat comes one interval after you set it. A loop also stops when a reply contains LOOP_COMPLETE on its own line, which its built-in instruction asks for once the task is finished.
How do I make it tell me only when something changes?
Hermes has no built-in “only on change” switch for these commands. You ask for it in the prompt:
- On Telegram, Discord, Slack and other messaging channels, a check whose whole reply is
[SILENT]orNO_REPLYis not sent to you. Upstream documents this for/heartbeat; the pinned source treats/loopchecks the same way, which we did not test. A reply that contains anything else is delivered. - In the terminal and the desktop or dashboard chat, the pinned source filters markers only in Bot Mode, so otherwise expect to see the marker.
- A marker at the end of a longer reply is removed and the rest is delivered, so ask for the marker alone.
- The built-in instructions don’t ask for silence.
/heartbeattells the model to “reply briefly that nothing has changed”, and/loopasks for a short report every time. Without your own[SILENT]instruction, you get a message on every check.
Hermes compares against what it can see in the conversation, so “same as last time” works because the earlier checks are there. Self-paced /loop uses the same idea to slow down: it compares each reply with the previous one, ignoring case, spacing, times, dates and durations.
What does each check cost?
Each check is an ordinary turn: Hermes sends your prompt as a message in the same conversation, the model reads the whole conversation so far, and it calls whatever tools it needs. The conversation grows with every check, so later checks cost more than early ones. Prompt caching helps less than you might expect: with Hermes’s default five-minute cache lifetime, checks five minutes or more apart usually find the cache expired. The long-conversation billing guide explains why. --until adds one judge call after every check.
Some numbers to work from: a 5-minute loop is 12 turns an hour. The default limit pauses a loop after 100 checks, and /loop resume continues. --times can stop it sooner but can’t raise that limit; change loops.max_ticks in the profile’s config.yaml (0 means no limit). A heartbeat has no limit and never stops itself. Check prices with your model provider.
What happens when I’m chatting, or Hermes restarts?
A check never interrupts a running turn, and a heartbeat waits for your queued messages. Missed heartbeat checks don’t pile up: you get one, not a backlog. A loop schedules its next check from the end of the previous one. An active /goal holds a loop back until the goal is done, paused or waiting.
Both are saved with the conversation. When the messaging gateway restarts, it restores active heartbeats and picks up active loops again without a new message from you. Starting a new conversation with /new, or switching away, drops that conversation’s heartbeat. After a restart, send /loop status or /heartbeat status. If the restart happened during a check, the pinned source can leave the loop waiting for that check’s reply until the next turn in the chat; any message from you, or /loop resume, gets it going again. We didn’t test a restart during a check.
When is a scheduled job the better tool?
| You want | Use |
|---|---|
| A short watch tied to what you’re doing now, with the conversation’s context | /loop |
| A standing reminder in this conversation, such as “every morning, check my inbox for the contract” | /heartbeat |
| A check that runs for days, many checks, or one that must not grow a conversation | A scheduled (cron) job, which runs each time in a fresh session |
| A check a script can read, where the model should run only when something changed | A scheduled job with a pre-check script: check without calling the model |
The last row is the true “tell me only when it changes” setup. When the script reports wakeAgent: false, or a monitor’s output hasn’t changed, no model turn runs at all.
What hosting changes
A watch on a messaging channel runs in the Hermes gateway, so it keeps checking only while the computer Hermes runs on stays on. If Hermes runs on your laptop, closing the laptop stops every watch (what closing the terminal stops). Vavio supplies a managed Hermes computer for $14.99 per month, with model usage billed separately, and keeps its gateway running. We haven’t checked whether the Vavio web chat runs /loop and /heartbeat the way upstream’s chats do; a messaging channel connected to your Hermes is the documented route. Hosting doesn’t change what each check costs. Compare the current hosting offer.
Sources
- Hermes v0.23.1 loops guide (/loop), checked 18 November 2026
- Hermes v0.23.1 heartbeat guide (/heartbeat), checked 18 November 2026
- Loop intervals, self-paced backoff and the 100-wakeup limit, checked 18 November 2026
- Heartbeat floor and prompt, checked 18 November 2026
- Silence markers the messaging gateway does not deliver, checked 18 November 2026
- Scheduled tasks (cron), checked 18 November 2026
- Vavio hosting offer, checked 18 November 2026
Written by Vavio with AI assistance and checked against pinned official Hermes v0.23.1 documentation and source. This is a source review, not a live watch, delivery or cost trial.