Tester Army e2e

What it is

e2e is an end-to-end testing framework for web and mobile apps, made by TesterArmy. A test can hold exact locator checks, plain-language goals that an agent carries out, or both. This listing covers the part that touches your coding agent. A skill teaches Claude Code, Codex or Cursor how e2e works. An MCP server gives the agent a live session on your app, and a trace page shows it what happened in a failed test. The agent walks a flow in the running app and writes each step into the test as it goes.

1. You describe a goal in plain words
2. An agent drives the app to reach it
3. A later check on the screen confirms the result
4. The run records what the agent did
5. Next runs replay it without a model until the app changes

Who it suits

Developers who ship a web or mobile app and want their coding agent to write and fix the UI tests, and to check its own UI changes in the running app. QA engineers with a Playwright, Cypress, Selenium, Detox or Maestro suite can move the suite over one slice at a time. Version 1.0 is not out yet, and the README says APIs and config can still change between minor releases.

Skip it if you need a settled 1.0 API, or if you work on Windows and will not use WSL: the quickstart says to run e2e inside WSL.

Where to use it

TesterArmy’s docs give full steps for three coding agents, and a folder for the skill that several more read.

• Claude Code: the skill, a plugin from the maker’s marketplace, and the MCP server
• Codex: the skill, the same plugin, and the MCP server in .codex/config.toml
• Cursor: the skill, a plugin in the local plugins folder, and the MCP server in .cursor/mcp.json
• Copilot, Gemini CLI, OpenCode and Zed: the skill in .agents/skills/, which the CLI reference says they read
• Grok Build: the docs say it reads the Claude Code plugin files

The tests run in a browser (Chromium, Firefox and WebKit through Playwright) or on iOS simulators, Android emulators and phones connected to your machine.

How to set it up

Get it saves a zip of Mindket’s setup notes in your downloads folder, with each agent’s steps from TesterArmy’s docs. The main route is the setup wizard, run in your app’s directory.

1. Run the install line. The wizard asks for an engine, a model provider, and where the skill and the MCP entry go, and shows every change before it writes anything.
2. Start your app’s dev server, or set its address and start command in e2e.config.ts.
3. Run the example test. It makes no model calls, so it needs no key.
4. Register the MCP server. The wizard offers to write .mcp.json for Claude Code and .cursor/mcp.json for Cursor. For Codex, add it by hand.
5. If you picked a subscription for agent steps, sign in, for example with npx e2e login openai.

npx e2e init
npx e2e run

The MCP server runs from the project’s own e2e install, so run the wizard and its install first. By hand, the line for Claude Code is below, and for Codex the server goes in .codex/config.toml.

claude mcp add --scope project e2e -- npx e2e mcp

[mcp_servers.e2e]
command = "npx"
args = ["e2e", "mcp"]

How to use it

Ask your agent in plain words. The MCP server page gives these prompts, and says that adding “headed” shows a local browser or simulator.

Add a Remember me checkbox to the login form. Check it in the app with e2e.
Write a test for checkout while you walk it in the app.
Reproduce issue #123 in the app and record a video.

While it iterates, the agent runs only the test it works on, by the line its test( opens on. A failed run writes a trace page under .e2e/results/ with the failing line, each step, the values an assertion read, console errors and failed requests. The agent reads that page and fixes every line it covers.

Example

You work on an app with a sign-in form and a checkout, and the MCP server is registered. You type:

Write a test for checkout while you walk it in the app.

The agent opens a session and takes each step on the live screen. Every action answers with the line that repeats it. The MCP page shows an answer of this shape:

Tapped #n31.
Test line: await screen.getByRole("button", "Sign in").tap();

The agent copies those lines into tests/checkout.e2e.ts. A goal such as “add the cheapest item” goes in as an agent.act step with an expect after it. You get a test file whose locator lines were checked on the live screen, and you can rerun it by line.

For newcomers and experts

If you are new. An end-to-end test uses your app the way a person does: it opens a page, types, clicks and checks what shows up. Written by hand, those tests are scripts of selectors and waits. With e2e you write a goal in words, and an agent does it. The docs say the agent adapts when the UI changes, so you fix fewer tests. Once a check on the screen passes, the run records what the agent did and replays it with no model calls. Start with the wizard in a small app. The example test needs no key.

For experts. A recording is written only when a later verification passes: a locator matcher, an engine assertion such as toHaveURL, locator.waitFor() or agent.assert. Reading a value with textContent() or expect.poll does not count. When a recorded control no longer matches, replay never guesses. It hands the rest of the step to the agent, which finishes from the current screen, and the run can save the new recording. Only agent.act replays. The docs say agent.assert, agent.waitFor and agent.extract always run live.

Renaming a test or rewording a goal makes its recording miss. The skill tells the agent to keep the cache on while it fixes a test and to turn retries off, since a retry runs every goal live. Once the test passes, it runs it one time with --no-cache to confirm that the live agent still reaches each goal.

Beyond tests, e2e explore gives the agent a goal in place of a test file and reports findings. A bug bash runs several of these at once, four at a time, each on one charter, and the agent keeps a finding only when a repro test fails. Each session spends model calls. The @e2e-dev/acp package can run agent steps on Claude Code or Codex over the Agent Client Protocol, with the agent’s own login, and keeps the agent to e2e’s step tools.

Good to know

Most of e2e runs on your machine, and a few paths reach out.

• Your permissions: tests, config, engines and app commands run with your operating-system permissions, unsandboxed. The docs say to run untrusted pull request code in an external sandbox with no secrets
• Usage data: the CLI sends anonymous usage data to PostHog at eu.i.posthog.com unless you opt out. The docs say it never includes test content, app content or credentials. Turn it off with npx e2e telemetry disable, E2E_TELEMETRY_DISABLED=1 or DO_NOT_TRACK=1. In CI, only the two variables turn it off
• Other calls: npx e2e feedback sends a report only when you run it, and the skill has the agent show the dry run first. A CLI command in a terminal asks registry.npmjs.org for a newer release at most once a day
• Model provider: agent steps send your instruction, the screen as an accessibility tree and the redacted URL to the model you configure. Screenshots mask secure fields only, so a secret typed into a plain field reaches the provider. Use test credentials
• Other sites: the docs say there is no origin allowlist, so the agent can follow the app to another site and use a step’s secrets there
• Hosted options: the Kernel, TestMu AI, BrowserStack, Limrun and EAS Simulators packages run browsers or devices on a service, so each needs that service’s key and sends traffic to it. The smol package runs machines on your computer, and its docs say they can reach any loopback port there
• Files: init edits package.json and .gitignore and writes a config, a test, the skill copy and the MCP entries. Cached recordings keep typed values except secret fills, so review committed ones like test code
• Real sites: the bug bash page says not to change data on a deployed site that other people use
• Removal: the docs give no uninstall steps. Mindket’s notes list what init wrote, so you can take it out by hand
• License: Apache-2.0. The LICENSE file names no holder, and the NOTICE file reads “Copyright 2026 TesterArmy, Inc.”