Hermes Agent for beginners: check five web pages and save a receipt

Five public pages went in. Two receipt files came out. In Hammer Automation's latest test, Hermes returned 5/5 PASS, and the observations in website-check-receipt.md matched website-check-evidence.json. The agent also read both files back before it finished.

This is a practical continuation of the beginner guide where you install Hermes Agent and complete your first verified task. This time, we leave installation and messaging channels alone. The goal is one real, bounded website job.

Who is this check for?

Imagine you look after a modest website and want to check the home page, an important article, the contact page, and a couple of machine-readable resources before publishing. You do not need to build a monitoring platform. You need a list of expected results and a receipt showing what the agent actually observed.

In one focused session, you will create:

• a checklist with five public addresses
• one direct Hermes Agent run
• a readable Markdown receipt
• the same observations in JSON
• your own spot check of at least one page

Before you start

Hermes should already respond in your terminal. Confirm that the terminal and file toolsets are available in your profile, and that curl, or an equivalent tool for HTTP GET, is installed.

Use public informational pages that you are allowed to check. This guide uses no sign-ins, cookies, forms, customer data, or keys. Request ordinary GET operations only and make no intentional changes to the site.

The Hermes documentation describes terminal as the tool for running commands and file as the tool for local files. The exact tools available depend on your installation and profile.

Source: Hermes Agent Tools Reference.

1. Create a dedicated working folder

On Linux, macOS, or WSL:

mkdir -p ~/hermes-site-check
cd ~/hermes-site-check

In Windows PowerShell:

$checkDir = Join-Path $HOME "hermes-site-check"
New-Item -ItemType Directory -Force $checkDir | Out-Null
Set-Location $checkDir

Starting in a dedicated folder makes the files created by the task easy to identify. The current Hermes CLI can also start in a chosen directory with --in, but cd is clear in both examples and was the route used in our test.

Source: Hermes Agent CLI commands.

2. Decide what counts as a pass

Create checklist.md. Each page needs:

• the requested URL
• the expected final HTTP status
• the expected final URL after redirects
• one or more public content markers

An acceptance criterion is an observable result that decides PASS or FAIL before the agent starts. A final URL is the address where the request lands after any redirects.

One item can look like this:

Requested URL: https://example.com/contact
Expected final status: 200
Expected final URL: https://example.com/contact
Expected marker: Contact us

The Hammer test used five real pages. They were the home page, both Swedish and English Hermes foundation articles, the contact page, and llms.txt. Pick equivalent pages on your own site. A clear heading or unique phrase is better than a common word that might appear in the footer.

3. Run the check with Hermes Agent

Save the checklist, then run:

hermes chat -Q -t terminal,file -q "Read checklist.md. For each public URL, make only an HTTP GET with redirects and a timeout. Use no credentials, cookies, request body, form submission, or state-changing HTTP method. Compare the observed final status, final URL, and every marker with the checklist. Save exactly two matching files: website-check-receipt.md and website-check-evidence.json. Use the same fields per route in both files, keep marker_checks as an array, and mark PASS only when every expected value matches. Read both files back, parse the JSON, confirm exactly five consistent routes, and finish with the saved paths and total pass count."

-t terminal,file selects only the two tool groups this task needs. -Q gives cleaner terminal output for scripts; it is not a security setting. The boundaries come from the checklist and prompt: public addresses, GET, no credentials or forms, and exactly two local output files.

Hermes v0.20.0 and the current CLI documentation support chat -q, -Q, -t, and --in. Check hermes chat --help if your version differs.

Source: Hermes Agent CLI commands.

4. Read the Markdown receipt

A useful receipt does not begin with a vague claim that everything looks good. It shows what was compared. Here is a shortened entry from Hammer's test:

requested_url: https://hammerautomation.ai/kontakt
expected_final_status: 200
observed_final_status: 200
expected_final_url: https://hammerautomation.ai/kontakt
observed_final_url: https://hammerautomation.ai/kontakt
expected marker: Kontakta oss
found: true
result: PASS

Look for three things on every row: the correct final status, the correct address after redirects, and the correct content marker. A page that returns 200 but lands on the wrong address should not slip through as a pass.

5. Compare the JSON file

JSON is not more truthful than Markdown. It is simply easier for the next workflow to read. An item might have this shape:

\{
  "requested_url": "https://example.com/contact",
  "expected_final_status": 200,
  "observed_final_status": 200,
  "expected_final_url": "https://example.com/contact",
  "observed_final_url": "https://example.com/contact",
  "marker_checks": [
    \{"expected": "Contact us", "found": true\}
  ],
  "result": "PASS"
\}

Reading the files back means the agent opens the saved files again and checks them instead of assuming the file write worked. Confirm that the JSON file contains five objects and that both files report the same total.

Our test receipt was unambiguous: five URLs, five final statuses, five final addresses, and all marker checks matched. Markdown and JSON both reported 5/5.

6. Perform your own spot check

Choose at least one row from the receipt and open the page yourself. Confirm that:

• you land on the reported final address
• the content marker is really there
• no sign-in, form submission, or account action occurred

Hammer repeated all five requests independently after the agent run. That check also returned 5/5. One spot check is often enough for a beginner routine; more consequential publishing workflows need more independent checks.

If a page fails

Read the reason before rerunning. A 301 or 308 can be an intentional redirect, but the final address still has to match the criterion. A missing marker can mean the page is wrong or the checklist is stale. Change the page or the criterion once you know which is true. Never turn FAIL into PASS because the result looks close enough.

What this check does not prove

This is a point-in-time check. It is not continuous monitoring, an SLA, a performance test, an accessibility audit, a visual regression test, or a test of purchases and forms. JavaScript-heavy and authenticated journeys need different tools and clearer access boundaries.

Hammer tested this exact workflow with a running Hermes agent on 12 August 2026. The agent receipt returned 5/5, and an independent check agreed. That proves the workflow worked at that time, not that the pages will always be available.

Once the manual check produces a trustworthy receipt, the next step could be a recurring run that alerts only on a real mismatch. Hammer Automation can help a team choose the right checkpoints and review steps in a Tool Forge implementation without starting with a large monitoring project.