Define a test once.
Run it forever.
Sisyphus clicks through your web, desktop and native apps the way a person would. Write the test in plain YAML, or let your coding agent write it. Every run is fast, repeatable and free of AI guesswork.
Tests you can read. Failures you can act on.
A test is a short list of what a person would do and see. Sisyphus finds elements by role and name, waits for them, and checks the result.
- One selector language. The browser, the native menu bar and the webview inside a desktop app, all in the same test.
- No sleeps, no flakes. Every check retries until it passes or times out. Flaky tests are flagged, not hidden.
- Failures that explain themselves. The step, an error code, the closest matching elements, a screenshot and the accessibility tree.
name: Add, complete and delete todos app: todo steps: - call: login - each: [Milk, Eggs, Bread] as: item do: - fill: textbox "New todo" = ${item} - press: Enter - expect: list "Todos" >> listitem count: 3 - check: checkbox "Complete Eggs" - expect: "#status" text: 2 of 3 left
✗ Add, complete and delete todos step 6 check: checkbox "Complete Eggs" ELEMENT_NOT_FOUND nearest: checkbox "Mark Eggs done" checkbox "Mark Milk done" screenshot: runs/…/006-fail.png tree: runs/…/006-fail.tree.yaml # agent: the label changed, update the step ✓ 1 passed · 1.2s
Your agent writes the tests. Sisyphus runs them.
Sisyphus ships an MCP server. Claude, Codex or any agent can explore your app step by step, save what it did as a test, and fix it when your UI changes.
- Exploring is writing. The agent drives the app one step at a time; the steps it took become the test.
- AI writes, it doesn't run. Runs use no model at all, so they're fast, free and give the same answer every time.
- Compact results. Small enough for an agent's context, precise enough to fix the test in one go.
The whole testing loop in one tool.
A CLI for CI, an MCP server for agents, and Sisyphus Studio for everyone else. Same tests, same engine.
Plain YAML, in your repo
Flows you reuse like functions, data rows, variables and secrets, if, repeat and each. Reviewed like any other code.
Record and inspect
Click through your app and Sisyphus writes the steps. Point at any element to get a selector that works.
A real debugger
Breakpoints, pause, step and try a step live. Every run keeps its screenshots and history.
Hybrid apps, finally
Test an Electron or Tauri app's native menus and shortcuts and the web content inside, in one test.
Load testing for free
Turn the same test into load: real browsers plus thousands of replayed users, with p95 and error-rate thresholds.
Made for CI
Parallel workers, retries, sharding, suites and JUnit output. Flaky and slow tests surface on their own.
Not another AI test runner.
Tools that put a model in every run are slow, costly and give different answers on different days. Sisyphus puts AI where it helps: writing and fixing tests.
| Web | Desktop & native | Runs | Who writes tests | |
|---|---|---|---|---|
| Playwright, Cypress | Yes | No | Deterministic | Developers, in code |
| Appium, WinAppDriver | Partly | One OS, a server to run | Deterministic | Developers, in code |
| AI-in-the-loop tools | Yes | Rarely | A model call per step | Prompts |
| Sisyphus | Yes | Yes, in the same test | Deterministic, no model | You or your agent, in YAML |
What you can test today.
Studio runs on macOS now. Windows and Linux native support is next.
| Web | Electron | Tauri | Native | |
|---|---|---|---|---|
| macOS | ✓ | ✓ | ✓ | ✓ |
| Windows | ✓ | ✓ | ✓ | Coming soon |
| Linux | ✓ | ✓ | Coming soon | Coming soon |
Push the boulder once.
We're letting teams in gradually during the alpha. Tell us what you'd like to test and we'll send you a build.
Request early access