Agent browser e2e
THE DEFAULT browser driver — invoke for ANY task driving a real browser, whether or not the user names it: E2E, QA, smoke, exploration, login-gated verification, resolving live selectors, plus team-workflow Phase 4 driving and Phase 4.5 exploration. Driver precedence: this skill FIRST → Playwright MCP only when its gate fails (say which condition) → claude-in-chrome only for the user's own logged-in Chrome profile. Reaching for a loaded browser MCP without running the gate is the known drift. Playwright still owns the committed `.spec.ts` suite — this drives, it is not a test framework. Headless login uses the encrypted Auth Vault, so the LLM never sees the password. Unattended runs (Phase 4/4.5, /team-run, scheduled) MUST have the dedicated E2E account + seeded test data in place BEFORE the first browser action — never invent credentials.From its SKILL.md
npx -y skills add JunjaK/ai-harness --skill agent-browser-e2eAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
2 things to look at
- no licenseNo license file was found in the repository. Code published without one is not open source by default, so using it at work is a question for whoever answers licensing questions where you are.
- 2 stars2 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
7.9 KB, as published. Nobody here has run it
agent-browser E2E
The default browser-driving layer (CLAUDE.md → "Browser Driving"). Whenever a real browser must be driven — E2E, QA, smoke, exploration, login-gated verification, resolving live selectors — come here FIRST, including inside team-workflow Phase 4 driving and Phase 4.5 exploration. You do not need the user to name it. agent-browser is a driver, not a test framework.
When to use / not use
- Use — any browser-driving task, requested or not: E2E / QA / smoke / exploration of a running app, Phase 4 driving, Phase 4.5 exploration, resolving real selectors for spec generation, and anything blocked by a headless login.
- Don't use as — the deterministic Playwright
.spec.tsregression suite. Driving is this skill's job; a flow that must regress forever still crystallizes to a Playwright spec (see below). - Don't reach past it — picking Playwright MCP or
claude-in-chromewithout first running the gate below is the drift this skill exists to stop.claude-in-chromeis for the user's own logged-in Chrome profile only.
Gate — check ONCE, then trust
Two conditions, evaluated deterministically at skill entry only. Cache the result for the rest of the task: do NOT re-probe per command, and do NOT use agent-browser doctor or agent-browser skills get core as the probe (those are heavy — that repeated firing is the waste this gate avoids).
- CLI present —
agent-browser --versionexits0(prints a version; no browser launch). - Skill available —
agent-browserappears in your available-skills list. This is a free observation (the harness already surfaces it) — issue no shell command for it.
| Outcome | Action |
|---|---|
| Both pass | agent-browser is the preferred driver. Proceed. |
| Either fails | State which condition failed in one line, then fall back to reference/e2e-testing.md (deterministic) / agentic-testing (exploratory) Playwright path. No silent skip. Fall back to Playwright, not to claude-in-chrome. |
After the gate passes, load the real usage guide once: agent-browser skills get core (+ agent-browser skills get dogfood for exploratory QA / bug-hunts). The guide is version-matched and is the SSOT for every command — do not re-load it per command, and do not duplicate its command reference here. Run agent-browser doctor only reactively — when a command fails unexpectedly — never as a gate.
Driving (CLI, native mode)
- Command surface (
open/snapshot -i/click @eN/fill/wait/screenshot/ …) lives in agent-browser'scoreskill — load it and follow it. - Core loop:
open <url>→snapshot -i→ act on@eNrefs → re-snapshot after any page change (refs go stale on navigation / re-render / dialog). - Isolation / parallel users: derive a stable id —
SESSION="$(agent-browser session id --scope worktree --prefix <task>)"— and pass--session "$SESSION".
Headless login (the core problem)
Headless mode cannot type a password interactively, and the LLM must never hold the secret.
Invariant — the LLM never receives the password in its context, command args, or output. The agent triggers a login by profile name; the secret is supplied out of band.
Default — Auth Vault
One-time save (the human runs this — it handles a real secret; the agent MUST NOT type the password):
echo "$PASSWORD" | agent-browser auth save <profile> \
--url https://app.example.com/login \
--username [email protected] \
--password-stdin
# add --username-selector / --password-selector / --submit-selector if the form is not auto-detected
Headless login (the agent runs this — no secret involved):
agent-browser auth login <profile> # waits for the form, fills, submits
Credentials are stored AES-256-GCM encrypted under ~/.agent-browser/. The agent references the profile name only.
Fallbacks (when the vault doesn't fit)
- External vault → credential plugin (team / CI; secret lives in real vault software). Gate the secret read explicitly:
agent-browser plugin add agent-browser-plugin-vault --name vault agent-browser --confirm-actions plugin:vault:credential.read \ auth login <profile> --credential-provider vault --item "My App" - Import state from a logged-in Chrome (best for 2FA / SSO — no password step at all). The human starts Chrome with
--remote-debugging-port=9222and logs in; then:agent-browser --auto-connect state save playwright/.auth/app.json agent-browser --state playwright/.auth/app.json open https://app.example.com/dashboard - Persistent profile / session restore (login survives across runs):
agent-browser --profile ~/.profiles/app open https://app.example.com # persists cookies/IndexedDB/SW SESSION="$(agent-browser session id --scope worktree --prefix app)" agent-browser --session "$SESSION" --restore open https://app.example.com # auto save/restore state
Security
- State / auth files contain live session tokens → keep them under the gitignored
playwright/.auth/(Playwright's own published convention — not_workspace/, which is for byproducts) and screenshots under_workspace/e2e/<run>/screenshots/(seee2e-testingArtifact Layout); delete when done. - Encrypt at rest:
export AGENT_BROWSER_ENCRYPTION_KEY=$(openssl rand -hex 32). --remote-debugging-portgrants full browser control on localhost — trusted machines only; close Chrome when done.- Starting the app server / tunnels is the human's job — surface a one-line "please start X" and don't script around it (harness operational discipline).
Verification (do not trust 200 / UI)
- A login is "successful" only when
agent-browser get urlis no longer the login page and a post-login element/text is present — not because a request returned 200. - An E2E goal is met only when the real outcome is observable (a persisted value survives reload; expected data renders). Apply
verification-loop's vacuity guard to any "it works" claim; express goals as outcomes, peragentic-testing.
Crystallizing to a regression test (optional)
On-demand exploration is not a permanent test. If a flow should regress forever, hand the verified path to e2e-testing and emit a Playwright .spec.ts following its conventions (getByRole > … > getByTestId; waitForResponse / waitFor, never waitForTimeout). agent-browser drives the exploration; Playwright owns the deterministic suite.
See also (link, do not duplicate)
- agent-browser CLI skills:
core(command SSOT),dogfood(exploratory QA),references/authentication.md(full auth patterns) reference/e2e-testing.md— deterministic Playwright.spec.tsconventions (fallback target + crystallization house style)skills/agentic-testing/SKILL.md— goal-verification concepts; the Phase 4.5 web Explorer drives through agent-browser when this gate passes, Playwright MCP otherwisereference/verification-loop.md— vacuity guard for "it works" claims