browser

Load access first when a task involves login, consent, MFA, CAPTCHA, or an account. Its API-first order decides whether a browser is needed.

Use three tools, cheapest first:

1. A saved script for a known flow. It runs without a model.
2. A Jev step inside that script, only when one element choice is uncertain or a step broke because the page changed. Code still checks the outcome.
3. agent-browser to explore a new flow, or when a script or Jev step gives up.

Save known flows

Once a flow works, save it as a Python Playwright script and run it without a model next time. A saved flow:

• takes its inputs as arguments, not edits to the file;
• gets credentials through the access fill pattern and never prints them;
• ends with an outcome check in code (expect(page.get_by_text("App created")).to_be_visible()) and exits non-zero when it fails;
• hands MFA, CAPTCHA and consent steps to the user, as access describes.

Jev for uncertain steps (optional)

Jev is TypeSafe's model for fast, structured decisions. Here a script asks it one closed question: which listed accessibility control performs this intent, or none. Act only on a confident answer, re-check that the control has not moved, then let code verify the outcome. Anything else goes to agent-browser. Never send secrets, codes or private values to Jev.

This repo does not ship a Jev step runner. Read references/jev-steps.md before building one; it covers the design, thresholds, prerequisites, and why an autonomous Jev browser agent is a different thing.

Explore and escalate with agent-browser

agent-browser is a browser CLI built for agents. Install it with npm install -g agent-browser, then agent-browser install to download Chrome for Testing.

agent-browser open https://example.com
agent-browser snapshot -i
agent-browser click @e3

Use the accessibility tree and role or label locators. Use screenshots only when the tree cannot answer the question. Keep any debugging endpoint inside scripts; never print one that carries a token. When exploration succeeds, save the steps as a script. When a saved flow breaks, use agent-browser to find what changed, then fix the script.

Never

• Touch a browser, profile, cookies, storage, process or debugging port that you did not start for this task.
• Expose a debugging endpoint or its token.
• Solve CAPTCHAs, bypass MFA, or use stealth or anti-detection tools.
• Use Playwright MCP on pages that receive secrets. Its accessibility snapshots have carried password values into model context (microsoft/playwright-mcp#1566).

Testing

For no-login tests, a local throwaway Chromium is fine. Keep its profile in your own scratch directory, bind its debugging port to 127.0.0.1, use a free port, and stop only the process you started. Keep Chromium's sandbox enabled. Run saved Playwright flows without a model and check the outcome in code. Use agent-browser for exploration.

Evidence

The October 2026 comparison behind these choices, on a synthetic site, is in benchmarks/README.md: data, method and limits. Read it only when you need to justify or revisit a choice.