e2e

e2e is an end-to-end runner from TesterArmy. The docs site is e2e.tester.army/docs. The npm package is e2e, bin e2e. License Apache-2.0. Contributors named in package.json: Oskar Kwaśniewski and Szymon Rybczak.

Checked 2026-10-07. The GitHub repo was created 2026-07-22. Default branch main. Language TypeScript. 6,950 stars, 310 forks, 73 open issues. Last push 2026-10-07. npm latest is 0.18.0, published 2026-10-06 (GitHub release tag e2e@0.18.0). packages/e2e/package.json on the tree read that day also says 0.18.0. main was pushed after that publish, so the branch can contain commits that are not in the tag. The README says the project is still on the way to 1.0, and that APIs and config can change between minor releases.

The same npm name already existed. Versions 0.0.1 (2014-12-16) through 0.0.8 are willscott/e2e, described on npm as a browser OpenPGP helper for encrypting, decrypting, signing, and verifying messages. TesterArmy versions start at 0.1.0 (2026-07-21, description then "Coming soon."). npm install e2e on 2026-10-07 resolves to 0.18.0. Maintainers of latest are okwasniewski and szymonrybczak.

How a test runs

A test can open the app, ask an agent to reach a goal, and then check the screen with locators. The README's example imports test and expect from e2e and uses app.open, agent.act, agent.assert, and screen.getByRole. Tests with no agent steps need no model. e2e init can be pointed at None for that path.

From the docs (cache and introduction, read 2026-10-07):

agent.act : Completes one natural-language goal. After a later check verifies it, the replay cache records the actions. The next run replays them with no model call until a control or the expected end state no longer matches, and then the agent takes over from the current screen.

agent.assert, agent.waitFor, agent.extract : Judge or read the screen. These always call the model. They are not replayed from the cache.

Locator assertions such as expect(screen.getByRole(...)) stay deterministic. The introduction says one test can mix goals and locators, and that each test runs once per target (a browser, or a simulator).

maxSteps and maxModelCalls default to 25 and go up to 100. agents.<name>.system is read by the act loop only. agents.<name>.context (at most 16 KiB) and a test's agentContext are also read by judges. The docs say nothing in system reaches a judge.

Packages

The repo is a pnpm workspace (packageManager pnpm@12.3.4 in the root package.json). Versions below are npm latest on 2026-10-07.

Example projects named in the README: Vite, Next.js, Expo, SwiftUI, Jetpack Compose, Kotlin Multiplatform, and Flutter.

Models

There is no default model and no shared API-key variable. Agent steps use the AI SDK. The built-in agent needs the ai package and a model that supports tool calls and images. Optional peer providers in e2e 0.18.0: @ai-sdk/anthropic, @ai-sdk/google, @ai-sdk/openai, @ai-sdk/openai-compatible, @ai-sdk/xai, and ai, all optional. The quickstart's sample config calls Vercel AI Gateway as gateway('openai/gpt-6-luna-fast'). That is the sample, not a model the runner selects on its own.

Subscriptions the docs tell you to sign in with, using the plan's own limits:

The subscriptions page says Claude subscriptions are not supported. Claude goes through an API provider, or through a model on a Copilot plan.

@e2e-dev/decision is a different path from the built-in agent. Its docs say a decision model answers one choice per action (which operation, and which element) from a semantic tree, and never generates free text. When the choice is type, a small language model writes the field value. The sample config uses @ai-sdk/typesafe-ai as typeSafeAi.decisionModel('jev-latest') plus an OpenRouter text model. The models page also names Clef as a decision model. This package is not the Decision 2.0 collection.

Install

Node engines, from packages/e2e/package.json and the quickstart: ^22.22.3 || >=24.8.0. The quickstart says to run inside WSL on Windows. Mobile setup also needs Xcode with an iOS simulator runtime, or the Android SDK with an emulator. npx agent-device doctor checks that machine.

npx e2e init
npx e2e run

pnpm dlx e2e init and bunx e2e init are the other forms in the quickstart. init asks for web or mobile, then a provider, and writes a config and an example test. The first example run in the quickstart has no model call. The coding-agent prompt on that page uses npx e2e init --yes, which it says writes a web config, an example test, the e2e skill, and MCP config, and installs nothing yet. npx e2e guide prints the skill.

The CLI sends anonymous usage telemetry by default: commands, engines, and where runs fail. The telemetry page says it does not send test content, app content, or credentials. Opt out with npx e2e telemetry disable (saved in ~/.config/e2e/telemetry.json), E2E_TELEMETRY_DISABLED=1, or DO_NOT_TRACK=1.

Security reports go to security@tester.army, not to a public issue. See SECURITY.md in the repo.

See also: Testing Tools · Decision 2.0 · The Future of Software Testing