Why Sisyphus
Sisyphus is an end-to-end testing tool for teams that ship web apps, desktop apps, or both. It’s built around one idea: a test should be written once and then run forever, cheaply, without anyone babysitting it.
What makes it different
Section titled “What makes it different”One format for every app
Section titled “One format for every app”Most tools test the web. A few test desktop apps. Almost none test an app that mixes both, such as an Electron or Tauri app with a native menu bar around web content. Sisyphus tests all of them with the same steps:
- click: menuitem "New Note" # the native menu bar- fill: textbox "Title" = Groceries # the web page inside the window- press: Mod+N # ⌘N on macOS, Ctrl+N elsewhere- expect: button "Groceries"You learn it once. Your web team and your desktop team write tests the same way, and one test can even move between apps: copy in one, paste in another, check the result.
Tests anyone can read
Section titled “Tests anyone can read”A Sisyphus test describes what a person does, in the words on the screen:
- fill: textbox "Email" = ada@example.com- click: button "Sign in"- expect: heading "Dashboard"No code, no CSS paths, no page objects. A product manager can review it. A support engineer can write one to reproduce a bug. When a test fails, everyone can see what it was trying to do.
Pointing at things by their role and visible name (button "Sign in") has a second benefit: it’s how assistive technology sees your app. Tests that find their way around easily mean an app that screen-reader users can use too, and Sisyphus warns you when they can’t.
Deterministic, with no AI at run time
Section titled “Deterministic, with no AI at run time”Some newer tools ask an AI model to interpret every step on every run. That’s slow, costs money per run, and gives different answers on different days.
Sisyphus runs your tests exactly as written, every time. No model, no per-run cost, no guessing. AI is welcome where it helps: writing and fixing tests. Then the test is plain YAML that runs deterministically forever.
No flakiness from waiting
Section titled “No flakiness from waiting”The most common cause of flaky tests is timing: the test clicked before the button appeared. In Sisyphus, every step waits for its target to exist and be ready, and every check keeps retrying until it passes or times out. You never write sleep. When a check does fail, it’s because the app really didn’t do it in time.
If a test fails and then passes on a retry, Sisyphus calls it flaky and keeps the failed attempts’ evidence, so flakiness gets fixed instead of hidden.
Failures that tell you what to do
Section titled “Failures that tell you what to do”When a step fails you get:
- the test, the step and its line number,
- an error code (
ELEMENT_NOT_FOUND,ASSERTION_FAILED, …), - the closest matching elements (“did you mean
button "Clear completed"?”), - a screenshot, and the list of everything that was on screen.
Most failures can be fixed from that alone, by a person or by an AI agent, without running the test again to see what happened. See Reading a failure.
Made for AI coding agents
Section titled “Made for AI coding agents”Sisyphus has a built-in MCP server, so coding agents (Claude Code, Cursor, and others) can use it directly. An agent can open your app, look at what’s on screen, try steps one at a time, and save what worked as a test. Then it runs the tests after every change it makes and reads compact results that point at what broke. See Tests with AI agents.
Your tests are your files
Section titled “Your tests are your files”Tests are plain text files in your repository, next to your code. They’re reviewed in pull requests, versioned with the code they test, and run the same on a laptop as in CI. There’s no cloud account to sign up for, no dashboard you have to log into, and nothing stops working if you stop paying for a seat.
What it gives your company
Section titled “What it gives your company”| Without end-to-end tests | With Sisyphus | |
|---|---|---|
| Releases | A manual test pass before each release, days of it | The suite runs on every change, in minutes |
| Regressions | Found by customers | Found on the pull request that caused them |
| Confidence to change code | Big refactors are avoided | Refactors are checked against what users do |
| Who can write tests | Engineers who know the framework | Anyone who can describe what a user does, and AI agents |
| Desktop and hybrid apps | Usually untested | Tested like the web app |
| Cost per run | People’s time | Effectively zero |
| Accessibility | Audited once in a while | Checked on every run |
And a few things beyond plain end-to-end tests, from the same test files:
- Load testing. Turn an end-to-end test into hundreds of virtual users and see what your backend does under pressure. See Load testing.
- Accessibility checks. Every run flags controls that screen readers can’t name or reach. See Accessibility checks.
- Layout checks. Resize to a phone and check nothing falls off the screen.
When Sisyphus is a good fit
Section titled “When Sisyphus is a good fit”- You ship a web app, a desktop app (macOS today), or an Electron or Tauri app.
- You want tests your whole team can read, not just one specialist.
- You use, or plan to use, AI coding agents and want them to check their own work.
- You’ve tried end-to-end testing before and gave up on flakiness.
Where it isn’t (yet)
Section titled “Where it isn’t (yet)”- Native Windows and Linux desktop apps. Web and Electron apps run on all three; native Windows and Linux app support is in progress. See Platform support.
- Visual pixel comparison. Sisyphus checks what’s on screen by meaning (text, roles, states, sizes), not by comparing screenshots pixel by pixel.
- Unit testing. Keep your unit tests. Sisyphus checks the product as a whole, on top of them.
Next: How Sisyphus works.
