Autonomous QA for web apps

Looks fine.
Look closer.

BugLupa is an AI testing platform we're building for web apps. You give it a URL and a test account for each role. Its agents drive real browsers through your app, try the flows they find, and report only what they can prove, with screenshots, traces, network logs and the exact steps to reproduce.

Meet the BugLupa spider. It follows your cursor and spins silk across this page, and any fly that lands in its web gets caught. BugLupa does the same for the bugs in your app. Press Esc to clear the web.

Illustration: a magnifying lens moves over an example order-management app. Under the lens, four problems appear that look fine at a glance: a viewer account can open an admin-only order by URL, the email field has no label, the helper text has 2.6 to 1 contrast, and the Save order button gets an HTTP 500 error without telling the user.

Authorization A viewer account opened order #1042 by URL. The page is meant for admins only.

One run, six kinds of checking

Teams usually spread this work across several tools and a lot of manual clicking. BugLupa is being built to run it together and keep the evidence in one place.

  • Functional

    Flows that actually finish

    Forms, records, search, filters, uploads and multi-step journeys, judged by what the page and the API really returned, not by whether a button was clicked.

  • Roles

    Permissions checked on the server

    Signs in as each role you provide and compares what each one can see and reach, including direct URLs and API calls. A hidden button doesn't count as protection.

  • Security

    Authorized, scoped and low-impact

    Passive checks by default: headers, cookies, CORS, sessions, exposed debug data. Active probes run only on targets you've verified, with rate limits and a stop button.

  • Accessibility

    WCAG 2.2 AA, labelled honestly

    Automated axe-core rules plus keyboard and focus checks. Anything that needs a person with a screen reader is marked as not tested, never as passed.

  • Performance

    Lab numbers with their conditions

    Lighthouse metrics, page weight, long tasks and caching, reported alongside the device and network profile they were measured on.

  • UI and UX

    Design review you can act on

    Key flows captured at seven screen widths in Chromium, Firefox and WebKit, reviewed for friction, consistency and empty or error states, with fixes a designer can pick up.

Every finding comes with proof

If BugLupa can't reproduce something, it won't call it confirmed. Each finding records what was done, as which role, in which browser, and what came back, so a developer can replay it instead of debating it.

  • Evidence before opinion. Playwright, axe-core and Lighthouse produce the evidence. AI plans the work and explains the results, but it doesn't get to invent them.
  • Your models, your choice. Use one model for everything, a different model for each agent, or automatic routing, with hosted APIs or local models.
  • Scope is enforced, not suggested. Allowed hosts, paths and actions are checked outside the model, so text on a page under test can't talk an agent past its limits.

Example of the format

BL-0142 Confirmed · High

Saving an order fails without telling the user

Route
/orders/new
Role
Manager
Browser
Chromium, 1440 px wide
Observed
POST /api/orders returned 500. The form showed no error and kept the entered data.
Expected
The order is saved, or a clear error explains what to do.
Evidence
Screenshot, trace, network log, console output
Reproduce
4 steps, with an exported Playwright test

Two ways to run it

The same testing engine powers both editions. Neither one is a trial of the other.

Planned first

Community

Free to self-host on your own server, with your own AI keys or local models. No usage caps, and no account with us.

Planned

Cloud

Run by Yarqat for teams: hosted browsers, shared projects and schedules, and the option to pay for AI usage with credits instead of your own keys.

Request early access