Skip to content

From a website address to a daily report.

Explore, test, run, heal, verify, report. One loop, on your schedule.

From address to report in seven steps.

  1. Step 1: Your URL

    Paste the address. Add login details if parts need one.

    Learn more

    Create a project with the website’s address. If parts of it are behind a login, add a login form (with an optional authenticator secret for 2FA), social login, session cookies, a bearer token or Basic Auth, then click “Test login” to prove it works before saving.

  2. Step 2: Crawl

    Maps pages, forms, buttons, links, lists and widgets.

    Learn more

    Up to 60 pages by default (adjustable). It closes cookie banners, follows single-page-app routes, and builds a map of pages, forms, buttons, links, lists and widgets. You can add PDF or Markdown documents for extra context. Detection is by page structure, never by site, vendor or builder name.

  3. Step 3: Tests

    Auto-Test writes one focused test per feature with Claude.

    Learn more

    One click: Page Load Checks, a Link Check and a Legal Links Check, one navigation test per menu link, one test per form, one per search box or AI assistant, and widget checks for tabs, accordions, filters, sorting and carousels. Add your own tests in plain English, if you like.

  4. Step 4: Run

    Site check and login first, then a real Chromium browser.

    Learn more

    Click “Run All Tests”, or let a schedule start it. Before any test, TestPilot checks the site is up and proves the login: it never runs tests logged out. Optional: watch the browser, record video, run a Lighthouse audit. An axe accessibility scan runs at the end of each web test.

  5. Step 5: Self-heal

    A button that moved is found again, within limits.

    Learn more

    A fingerprint match on the live page first, then Claude reads the page’s HTML, within 25 seconds per step. A fix is saved back to the test with version history.

  6. Step 6: Verify

    Evidence decides each check. Vision only where rules can’t.

    Learn more

    Network answers, page changes and the address the browser ended on decide first. Claude looks at the screen only where rules can’t decide, and always before passing something the site produced.

  7. Step 7: Report

    Real site bugs first, a screenshot for every step.

    Learn more

    Then warnings, blocked tests, tests needing maintenance, tests not run, and passes. Every step and every result has a one-screen screenshot, and failures get a plain root-cause note.

Four rules that keep results honest.

A test only passes when its main check really ran.

When a button moves, the test finds it again.

A fingerprint match first, then Claude reads the page. Nothing within 25 seconds? The step fails honestly.

Learn more

A fix is saved back to the test with version history and counted on the run. Never healed: counts, waits, search and filter checks, steps in embedded frames, and elements the crawl saw only on another page.

  • First a fast fingerprint match on the live page, then Claude reading the page, within 25 seconds per step.
  • A healed selector is saved back to the test with version history, and the run counts healed steps.
  • Never healed: count checks, wait steps, search and filter checks, steps inside embedded frames, and elements the crawl only saw on another page.
  • IDs that UI libraries regenerate on every page load are never used as selectors.
Button not foundIt moved
Fingerprint match1 · fast
Claude reads the page2 · if needed
Found: fix saved
25 s up: fails honestly

Evidence first. Vision only where rules can't decide.

Network answers, page changes and the final address come first. Claude looks at the screen only when they can’t decide.

Learn more

Each check reads the responses and page changes caused by the test’s own action: a 500 is a failure, whatever the page shows. Claude always looks before passing something the site produced, like a summary or a chat reply. The step log records what decided each check.

  • Rule-based checks come first: page loads, address, visible, text, count, enabled, search, filter and sort.
  • Network and page evidence is measured from the test’s own action, not from whatever the page did earlier.
  • Claude’s vision decides only where rules can’t, such as a visual check or a coloured badge.
  • Anything the site produces (a summary, a generated result, an assistant’s reply) never passes on network evidence alone: it has to be seen.
  • The step log records what decided each check. A false pass is treated as worse than a false fail; final checks are never optional.
  1. Network1
  2. Page2
  3. Address3
  4. Vision4

Site down? One clear message, not fifty failures.

Checked before the login and any test, then again 30 seconds later. Still down: the run stops.

Learn more

No answer (DNS, refused, timeout, dropped) or a 5xx page counts as down; a CAPTCHA or bot check doesn’t. Its tests are marked Not run, not failed, and a scheduled run emails a short “Site down” alert. If TestPilot’s own connection is the problem, nothing is blamed on your site and no email goes out.

  • Before the login and before any test, it checks the site answers.
  • No answer, or a server error page, is checked again 30 seconds later.
  • Still down: the run stops, its tests are marked Not run, and a scheduled run emails a short “Site down” alert instead of a report.
  • A CAPTCHA or bot check is not an outage. If TestPilot’s own internet connection is down, nothing is blamed on your site and no email goes out.
Does it answer?Check 1
Check againCheck 2
Up: tests run
Still down: run stops, tests Not run

Real submissions, with test data you can spot.

Forms are filled and submitted for real, with values your team can recognise and filter out.

Learn more

Emails go to example.com, which can’t receive mail, and are new every run. Phones get +15550100199, or 9876543210 where a 10-digit number is required. Messages say “Automated test by TestPilot - please ignore”. Honeypots and newsletter opt-ins are left alone; values from your Data tab are used exactly as given. Each run submits each tested form once, and TestPilot doesn’t delete those entries afterwards.

  • Every visible field is filled, required or not, including forms without a form tag, forms in frames, custom dropdowns and file uploads.
  • Hidden spam traps (honeypots) and newsletter opt-ins are left alone.
  • A form that answers “already subscribed” passes with a warning, because the form works.
  • Forms that create something get a fresh run tag in their name or email, so day two isn’t “already exists”.
  • No negative tests: it doesn’t submit empty or invalid values on purpose.
  • Each run submits each tested form once, so a daily schedule leaves one labelled entry per form per day in your inbox, CRM or database. TestPilot doesn’t delete them; filter on testpilot.*@example.com or the “please ignore” message.
  • Auto-Test doesn’t skip order, booking or support forms on its own. To keep one out, schedule only the tests you choose, or delete that form’s test.
Your form, during a run
Emailtestpilot.<run tag>@example.com
Phone+15550100199
MessageAutomated test by TestPilot - please ignore
FileA real, readable test file
SubmitOnce per run

Then it runs on its own.

Schedules run the same tests every day, the report lands in your inbox, and the health report shows the trend.

What you need to run it.

TestPilot is self-hosted today. This is the whole list.

A machine you control

TestPilot and its scheduler run there, so it has to be switched on at your schedule times.

Node.js

TestPilot is a Next.js app. Its Chromium browser is installed with its packages (npm install).

Claude Code, signed in

AI runs through the Claude Code command-line tool on that machine, under its Claude account, so it uses that account’s plan.

An SMTP account, for email

Only for emailed reports from scheduled runs. A Gmail app password by default, set in .env.local.

What it costs to run

That machine, the Claude plan behind its Claude Code login (every AI call counts against it), and an SMTP account if you want emails. TestPilot's own pricing hasn't been decided. Where your data goes is on the security page.

See TestPilot on your own website.

Book a demo to see a crawl, the tests it writes, a run and its report, on a website you choose.