e2e: the open-source AI testing framework

Selector-based tests break when the UI changes. With e2e you write the goal ("upgrade the workspace to the Pro plan") and an agent clicks through the app. Deterministic checks still run as usual.

How it works

1. 1Describe the goal. agent.act('upgrade the workspace to the Pro plan') drives the app; agent.assert() checks the result in plain English.
2. 2Mix in real assertions. Locators and expect() work in the same test, so the parts you want deterministic stay deterministic.
3. 3Replay cache. Once a later check confirms an agent step worked, its actions are recorded. The next run replays that step with no model call. agent.assert always runs live.
4. 4When the UI changes. If a recorded control or end state no longer matches, the agent takes over from the current screen.
5. 5Web, iOS, Android, CI. Playwright for web (Chromium, Firefox, WebKit); iOS simulators and Android emulators for mobile; hosted options via Kernel and EAS Simulators; runs locally or in CI.
6. 6Your own model. On your machine, sign in with ChatGPT Plus or Pro, GitHub Copilot, or SuperGrok / X Premium+. In CI, use an API key (e.g. Vercel AI Gateway). Claude subscriptions aren't supported.

Get started · Node.js 22.12+

Good to know

• Still before 1.0: APIs and config can change between minor releases.
• The CLI sends anonymous usage telemetry by default; turn it off with npx e2e telemetry disable.
• In our test: the first run's agent step took 4 model calls; the replay took none (the live agent.assert took 1).
• An independent guide by @copilot_shogo. Not affiliated with TesterArmy.