Can Hermes watch a page or run a check without calling the model every time?
Yes. Hermes Agent can run a scheduled script with no model call, or run a pre-check script that wakes the agent only when there is something to interpret. Asking the model to reply with [SILENT] is different: that suppresses delivery after the agent runs, so it does not avoid the model call. A delivery to another agent can still trigger that recipient’s model turn. On Vavio, hosting and any model usage remain separate charges.
This follows the official Hermes v0.23.1 documentation, commit 3b42157, checked on 15 November 2026. We have not tested a page watcher, alert delivery or savings on a Vavio computer. The examples explain upstream Hermes controls, not a preconfigured monitoring service.
Which kind of check needs a model?
Use a script when the rule and the message can be determined mechanically. Use the agent when the result needs interpretation.
| Job | Documented approach | Does this job’s execution call the model? |
|---|---|---|
| Check a threshold and report its value | Script-only job with no_agent=True |
No. |
| Check for new data, then summarize it | Pre-check script with wakeAgent |
Only when the script allows the agent to run. |
| Ask the agent to judge whether a change matters | Ordinary agent job; optionally suppress quiet results with [SILENT] |
Yes, including quiet results. |
A page-change watch can fit either of the first two approaches: a script fetches and compares the relevant data, then either writes a complete alert or passes the changed data to the agent. The script still needs to handle the site’s access rules, authentication and failures. Calling it a watch does not make a blocked page readable.
What does a script-only job do?
In no-agent mode, the script is the job. Hermes runs it on the schedule and uses its output directly, without asking a language model to rewrite or judge it.
The cron guide defines these outcomes:
- A successful script with non-empty standard output has that trimmed output delivered verbatim.
- A successful script with empty standard output produces a silent tick.
- A non-zero exit or timeout produces an error alert under the documented behavior.
- A final standard-output line containing
{"wakeAgent": false}also makes the tick silent.
The job makes no inference call, so there are no model tokens for that scheduled execution. Delivery can trigger separate inference: the guide says a bot-chat delivery costs the receiving bot a full agent turn. Skipping the sending job’s model does not skip that recipient’s model. The computer, network and any external API the script uses can still have costs. An assistant may also use model calls while helping you write and set up the script.
In a shell with the appropriate Hermes profile available, the upstream CLI shape is:
hermes cron create "every 5m" \
--no-agent \
--script check-feed.py \
--name "feed-check"
This assumes you have already written check-feed.py inside the active Hermes home’s scripts/ directory. Add a delivery target only after it is configured and you have checked where output should go. The example does not prove a particular messaging integration on Vavio.
A script argument is a file, not an inline shell command. Hermes requires the resolved path to stay inside $HERMES_HOME/scripts/. .sh and .bash files run with Bash; other extensions use the current Python interpreter unless a supported user-managed interpreter is configured. Scripts do not automatically inherit Hermes-managed model-provider credentials.
Because standard output can become the delivered message, avoid printing secrets or full private API responses. A script should also distinguish an empty valid result from an error: an authentication failure should not look like “nothing changed.”
How do I call the model only when data changes?
Attach a pre-check script to an agent job and let it decide whether inference is needed. The documented gate is a final line of JSON on standard output:
{"wakeAgent": false}
That skips the agent for the tick. If wakeAgent is omitted, the default is true: the agent runs normally. The guide also shows passing information to the agent when waking it:
{"wakeAgent": true, "context": {"new_items": 3}}
For a change watch, the script can fetch the current value, compare it with its saved baseline and return false when unchanged. On a change, it can pass the new information for a summary. You still have to define what counts as a relevant change and how baseline state survives failures. A whole-page comparison may wake the agent for a timestamp, rotating advertisement or login page.
This is a documented design pattern, not a tested watcher recipe. For important checks, exercise unchanged data, a relevant change, an HTTP error and an expired login before depending on it. Do not replace the last good baseline with an error page.
Does SILENT avoid the model bill?
No. [SILENT] is a delivery control for the agent’s final response. The agent has already run to produce that response.
If a successful agent reply is [SILENT], starts with [SILENT], or has a silence marker as its own first or last line, Hermes saves the output locally but sends no message to the delivery target. A reply that only mentions [SILENT] mid-sentence is still delivered. Failed jobs still deliver regardless of that marker. It is useful when judgment is necessary on every check but only some results deserve your attention. It is not a substitute for a pre-check when the goal is avoiding unnecessary inference.
What still has to stay running?
These jobs use Hermes’s scheduler. A script-only check does not become an independent system monitor just because it skips the model. If the Hermes gateway is down, an in-gateway watchdog cannot reliably be the thing that tells you it is down. The script-only guide recommends an independent OS-level check for that case.
For ordinary scheduled work, does closing the terminal stop Hermes? explains the difference between an interactive process and a supervised gateway. A running computer alone does not prove your script completed or its result reached you.
Where should I start?
Pick one simple check whose output you can verify. If the script can produce the complete result, use no-agent mode. If you need a summary only when new data arrives, use a pre-check gate. If every result needs judgment, keep the agent and use quiet delivery where appropriate.
When the agent does run, you can also give scheduled jobs a different model from chat. That changes the inference route, while a pre-check decides whether to use inference at all.
Vavio provides a managed remote computer for Hermes at $14.99 per month, with model usage billed separately. It can keep the computer available away from your laptop; you still need to configure and verify the particular check. Compare the current hosting offer, or use the same upstream controls on your own server if you prefer to manage it yourself.
Sources
- Which replies count as silent (source), checked 16 November 2026
- Hermes v0.23.1 script-only cron jobs, checked 15 November 2026
- Cron no-agent behavior and script constraints, checked 15 November 2026
- Pre-check scripts and wakeAgent gates, checked 15 November 2026
- SILENT delivery suppression, checked 15 November 2026
- Bot-chat delivery triggers a recipient agent turn, checked 15 November 2026
- Vavio hosting offer, checked 15 November 2026
Written by Vavio with AI assistance and checked against pinned official Hermes v0.23.1 documentation and source. This is a documentation and source review, not a reliability, delivery or cost trial.