Coming soon · Web apps on Chromium

The Quality Engineering computer

Give QEvin a test app. It explores the product, shows the evidence for everything it learns, maps journeys into a knowledge graph and writes Playwright tests your team reviews.

Runs on your desktop with your own Claude login. Your test app, your rules.

mission m_0142Running
1
pages mapped
1
facts with evidence
4
locators
0
tests validated
Labs Supply · demo store
sample mission
Labs Supply products page Stoneware mug product page Stoneware mug added to the cart Shopping cart with the Stoneware mug Checkout asks guests to log in
labs-supply.test/demo/
Step 01· 2 of 25 actions · 0 min 18 sRunning
Opened /demo/
Observed

QEvin drives a real Chromium window and asks before anything sensitive.

1

path, done well: Chromium, Playwright and TypeScript.

100%

of facts link to a screenshot, a DOM snapshot or a replay.

0

secrets in prompts, logs, screenshots or scripts.

4

budgets on every mission: time, actions, tokens, retries.

01Evidence

Every fact shows its work

QEvin never asks you to trust a sentence. Each fact it learns is pinned to a screenshot, a DOM snapshot and a replay, and it carries an honest status.

screenshot · s_0931.jpg
fact f_2207 · Cartknowledge v13
“Proceed to checkout” leads to a login prompt for guests
Inferredstep 08
Looked like the start of checkout. Not proven, so not shown as fact.
Observedstep 09
Clicked once with your approval and saw the login dialog.
Validatedreplay r_0412
Replayed in a clean browser. Passed 3 of 3.
Human-confirmedby Priya
The strongest status a fact can have.
Later it can turnStaleorConflictedand it says so.
dom snapshot · d_0931
<button class="btn" id="checkout" type="button"> Proceed to checkout </button>
getByRole('button', { name: 'Proceed to checkout' })3 / 3
replay · r_0412
Run 1 · clean profilepassed
Run 2passed
Run 3passed
provenance
Missionm_0142
App build2026.10.27 · staging
Knowledgev13
Confirmed byPPriya · QA lead
02Guardrails

The model proposes. Code decides.

Every action passes a tool gateway before it runs. Policy lives in code, so text hidden in a page or a ticket can't talk its way past it.

Proposed by the modelTool gateway · policy.tsOutcome
click · Add to cart
Allowed · ran
click · Proceed to checkout
Needs approvalAllow once
click · Delete account
Blocked · destructive
read · document.cookie
Blocked · secrets are never read
Actions12 of 25
Time4 min 10 s of 15 min
Retries2 of 3

Budgets on every mission

Time, actions, tokens and retries. When one runs out, QEvin stops safely and keeps what it found.

Close the app. The mission keeps going.

Missions are durable jobs with explicit states. Reopen QEvin and it is right where it was.

13 fill #email tester@example.test
14 fill #password •••••••• [redacted]
15 cookies never read

Secrets stay out

Never in prompts, graph nodes, logs, screenshots or scripts. Credentials live in your OS keychain.

Permitted test apps only No production exploration Page text is data, never instructions People approve publishing
03Replay

Explore once. Replay forever.

QEvin explores with computer use, then turns what it proved into deterministic Playwright tests. No model in the loop when your suite runs.

COMPUTER USE · MISSION m_0142
01Opened /demo/
02Clicked “View product” on Stoneware mug
03Clicked “Add to cart”, then “View cart”
04Clicked “Proceed to checkout” (allowed once)
05Saw the login prompt for guests
PLAYWRIGHT · checkout.spec.tsValidated
test('guest is asked to log in at checkout', async ({ page }) => {
  await page.goto('/demo/');
  await page.getByRole('link', { name: 'View product' }).first().click();
  await page.getByRole('button', { name: 'Add to cart' }).click();
  await page.getByRole('link', { name: 'View cart' }).click();
  await page.getByRole('button', { name: 'Proceed to checkout' }).click();
  await expect(page.getByRole('dialog', { name: 'Log in to check out' })).toBeVisible();
});

Deterministic replay

Plain TypeScript on Chromium. Runs the same way in CI every time.

Validated before you see it

Each script is replayed. A failed or inconclusive run is shown as exactly that.

Readable locators

Role and text first, chosen from the locators the graph has seen work.

You approve the merge

QEvin opens a review. It never publishes code on its own.

04Knowledge

A map of your app that remembers

Pages, journeys, locators and rules, each with its evidence and version. Not a transcript, not a vector blob. It gets sharper with every mission.

Products page
Products
Validated
Product detail page
Product detail
Validated · Add to cart
Cart page
Cart
Validated · 3 facts
Checkout login prompt
Checkout · login
Observed · new in v13
Log in
Inferred · not visited

Versioned

Every mission writes a new version. Nothing is overwritten.

Diffable

See what changed between builds: new pages, moved buttons, broken journeys.

Honest about age

Facts from an older build turn Stale. Disagreements show as Conflicted.

05Publish

Publish it for the team

A mission ends in a knowledge version your team can review, not a chat log. Approve it, and the graph and tests become shared.

1
Review the diff
New, changed and stale facts, each with its evidence.
2
Confirm what matters
A person's confirmation is the strongest status a fact can have.
3
Ship the tests
Validated scripts go to your repo as a pull request you merge.
Knowledge v12 → v13Waiting for you
+14 facts3 changed2 stale1 conflicted
+ page Checkout prompt “Log in to check out” + link Cart → Checkout “Proceed to checkout” ~ stale header cart count ! conflict Log in is a page or a dialog?
Review request2 of 3 approved
PPriya · QA leadApproved
AArjun · SDETApproved
MMeera · ProductWaiting
Playwright · CI3 passed
checkout.spec.tspassed · 4.1 s
cart.spec.tspassed · 2.8 s
product-detail.spec.tspassed · 1.9 s

Let QEvin explore your test app

We are finishing the first release for web apps.

macOS first. Chromium and Playwright with TypeScript.