Get Vargus to verify it.

The verification agent for teams shipping with coding agents. Tests in markdown, run from your terminal and CI, against the real application: web, API, CLI and desktop.

app.vargus.ai/console
Vargus dashboard showing test run metrics
Claude
Open models
Playwright
Linux desktop
macOS desktop
GitHub Actions

Measured, not promised

Numbers from Vargus's own runs, as of 28 October 2026. Every one has a build and a repository behind it.

0

false accusations

across 465 tests on a production web app, three executions, every confirmed bug reproduced before it was reported.

3

real bugs confirmed on macOS

two in CotEditor (7.0.7, 7.1.0) and one in IINA (1.4.4), each caught on the official pre-fix build and passed on the fix.

88%

of the top 500 Flathub apps

expose an interface Vargus can read and drive. 16 desktop toolkits on Linux, 24 apps on macOS through the accessibility tree.

$0.16

per web test

about half a minute each. A planted desktop bug caught by Haiku for $0.02. Every run shows its cost before and after.

How a verdict is earned

1.

Tests in markdown that describe intent, not selectors. Vargus drafts a starting suite from your code; you keep what is right.

2.

One command, on your machine or in CI. The agent drives the real application: web, API, CLI or desktop.

3.

A reported bug is replayed on a fresh copy before you see it. Unconfirmed findings are kept, never turned into blame.

4.

Exit codes, JUnit XML, GitHub Actions annotations, and the cost of every test beside its verdict.

Write tests in plain markdown

Describe what should happen. The agent works out how to do it in the real application.

app.vargus.ai/console/suites/auth-flows
Vargus suite detail showing performance and flaky test history

No boilerplate

Tests describe intent, not implementation details. No fragile setup to maintain.

No abstraction layers

No wrappers or helpers. Write what should happen in natural language.

AI handles execution

The agent works out how to reach each step in the running application and verifies it there.

Real environment testing

Every test runs against your actual application, started with its own state, never a mock or a simulation.

Desktop applications too

Cloud QA services stop at the browser. Vargus tests the desktop app you ship, on the machine it ships to.

Driven through the accessibility tree

Vargus reads the window the way a screen reader does and acts on named controls. No screenshots unless the tree is missing, and then it says so rather than guessing.

Launched with its own state

Every suite and every replay starts the application fresh, with isolated state, and stops only what it started.

A bug is replayed before it is reported

Every action is recorded and timed. A reported bug is confirmed by replaying those actions on a fresh copy, without the agent in the loop.

Linux and macOS

GTK, Qt, Electron, Flutter, wxWidgets, Swing, Tk and more on Linux; AppKit, SwiftUI, Qt, Electron and Apple’s own apps on macOS. Windows is not yet measured.

Tests without the boilerplate

Same test. Fraction of the code. Zero selectors to maintain.

app.vargus.ai/console/runs/run-9438
Vargus run detail showing suite progress and logs
app.vargus.ai/console/runs/run-9436
Vargus run detail of a desktop application suite with one bug confirmed by replay

Built for production workflows

Cost controls, parallelism, CI integration and analytics, all local to your repo.

Budget Controls

A cost estimate before the run, cost per test during it, and a budget cap you set. Pay your own model provider, or run open models.

Parallel Execution

Suites run concurrently with hardware-aware limits in the browser, and one at a time on the desktop, where they share the machine.

CI/CD Integration

One command for CI. Exit codes, JUnit XML and GitHub Actions annotations built in. Stale suites are refreshed only when you say so.

Analytics

Pass rates, cost trends, flaky tests and a diff against the last run. A run that verified nothing is never shown green.

What Vargus refuses to do

A verifier is only as good as what it will not let itself get away with.

Scripted clicks and page edits

A click injected by script raises none of the events a real one does. Vargus refuses it and clicks a visible point of the element with real pointer events, naming whatever covers it.

Passing a test it never ran

A PASS reported without touching the running application is downgraded. A run in which nothing was verified is not a success.

Blaming the software on one look

An accusation is replayed or re-run before it is printed. Confirmed is a bug. Unconfirmed keeps the finding and withholds the blame.

Falling back to screenshots in silence

If a desktop application’s accessibility tree is unreadable, the run refuses rather than quietly switching to pixels.

Spending without saying so

A test that sends mail, charges a card or destroys data declares it, and skips at $0 unless the run sanctions that effect.

Surprising you with the bill

Cost is estimated before the run, tracked per test during it, and capped by a budget you set.

Join the waitlist

Be first to try Vargus when we launch.

We respect your privacy completely. No spam and no third party sales. Unsubscribe at any time.