How do you monitor a website for changes nobody deployed?
Last updated: 22 November 2026
Keep a snapshot of the site as it should be, and compare the live site against that snapshot on a schedule, page by page, on three checks: every page still exists, its content is unchanged, and its screenshot looks the same. A page that changed without a deploy, such as an edit in the CMS, a plugin update or a broken embed, then fails the next scheduled run.
Who runs the schedule
Your own scheduler starts each run and reports a failed one: a GitHub Actions schedule, a cron entry or the host of an agent. DiffiCat runs no schedule and sends no alert; it compares when a run asks it to and returns the result, and the scheduler’s own failure notice tells you.
The snapshot and its id
Take the snapshot in the web app while the site is as it should be, and switch its Keep on. A snapshot saved from a comparison run can be archived to make room for newer runs, and an archived snapshot fails every scheduled run until you restore it from the web app’s Archive. The id is the part of the snapshot’s address that starts with snap-. A run checks only the pages the snapshot holds. After a change you made on purpose, take a new snapshot and replace the id, since the old one keeps reporting that change.
Which key a run needs
A run needs the key of the account that holds the snapshot, copied on the account page and read from your scheduler’s secret store or the DIFFICAT_KEY environment variable. Any other key, a mistyped or replaced one included, gets the same not-found answer as a wrong id. So a run that exits 3 with not found has one of three causes: a wrong id, an archived snapshot, or a key that is not the key of the account holding the snapshot. A prepaid key is debited for each run; a Pro account’s key draws on its monthly allowance. Either way a run draws one captured page per page of the snapshot. Keys and setup are described on the agents page.
What a schedule draws
The ceiling in each snippet is your own figure: a run priced above it is refused before it starts. It bounds one run, not the month; how often the schedule runs sets what a month draws. A settled charge is final. The command sends no idempotency key, so a step your scheduler retries is a new run and is charged again. Before you schedule anything, run the shell line once with --dry-run: it prints the quote and charges nothing, and it fails with not found when the id or the key is wrong.
GitHub Actions
name: Scheduled site check
on:
schedule:
- cron: '0 6 * * *'
workflow_dispatch:
permissions:
contents: read
jobs:
check:
runs-on: ubuntu-latest
steps:
- uses: difficat/difficat@v1
with:
snapshot: snap-00000000-0000-0000-0000-000000000000
target: https://www.example.com
max-price: 2
key: ${{ secrets.DIFFICAT_KEY }} The workflow runs every morning and from the Actions tab. The step fails when a page fails, and the job summary names the failed pages. Store the key as the repository secret DIFFICAT_KEY.
A cron entry
npx difficat compare --snapshot snap-00000000-0000-0000-0000-000000000000 https://www.example.com --max-price 2 Exit codes: 0 every page passed, 1 a page failed, 2 payment required, 3 an error. A non-zero exit is what your scheduler reports. The command reads the key from the DIFFICAT_KEY environment variable; set it in the cron environment, not on the command line. npx fetches the newest published release when none is installed; write difficat@ followed by a version to pin one.
An agent on a schedule
compare_snapshot({
"snapshotId": "snap-00000000-0000-0000-0000-000000000000",
"targetUrl": "https://www.example.com",
"maxPriceUsd": 2
}) The compare_snapshot tool is served at https://app.difficat.com/mcp?surface=full and by the plugin, not at the default address. The key is sent by the MCP client as a request header in its configuration and is never passed as a tool argument. An agent that retries the call passes the same idempotencyKey again, so the run is not charged twice.
What this does not do
- The live site must resolve on the public internet. Localhost, loopback and private-network addresses are refused.
- Pages are rendered in headless Chromium only, at the snapshot’s width and check mode.
- Content that differs between two loads of the same page, such as a rotating banner or a timestamp, is reported as a change, so that page fails on every run.
- The service refuses a snapshot set of more than 200 pages on every run, and the run exits 3, so check the page count when you take the snapshot and keep a smaller set.
- An intended change is reported like any other, so a person or an agent still decides.
Checking a deploy instead? The guide check every deploy against production compares a preview with production from the pipeline that deploys.
Try it on your own site
Enter your site in the try-it form to see the report before you schedule a check.