dev-browser

Do every browser task in Chrome AI, the user's dedicated Chrome window (profile ~/.chrome-ai-profile, port 9333) where they are already signed in to their tools. Start it first (it does nothing if it is already open), then connect to it on every run:

chrome-ai start
dev-browser --connect http://127.0.0.1:9333 <<'JS'
const page = await browser.getPage("stripe-sales");
await page.goto("https://dashboard.stripe.com");
console.log(JSON.stringify({ title: await page.title(), url: page.url() }));
JS

If chrome-ai or dev-browser is missing, stop and tell the user to run /setup-dev-browser; never fall back to a standalone Playwright script or the user's own Chrome. If CHROME_AI_PORT is set, connect to that port instead of 9333 (chrome-ai url prints the address).

• Give each task its own page name (browser.getPage("lumail-campaign")) so parallel agents never share a tab, reuse that name across runs to keep the page state, and close it with browser.closePage(name) when the task is done. Never close or navigate a tab you did not open.
• Write small scripts that do one thing and end by logging the state as JSON (url, title, what you found). Scripts run in a sandbox, not Node: no require, fs, fetch or process.
• Before clicking, read the page with await page.snapshotForAI() and target elements with page.getByRole, getByText or getByLabel. Open references/api-reference.md before using a method you are not sure exists.
• Files only go to ~/.dev-browser/tmp/: pass a bare file name to saveScreenshot or writeFile and use the path it returns. Look at the screenshot before saying a page looks right.
• If a site asks to sign in, stop and ask the user to sign in in the Chrome AI window, then continue; the session stays saved. Never type passwords, codes or card numbers, and never solve a captcha.
• Ask the user before any click that sends, publishes, pays, deletes or accepts terms.
• For a clean signed-out check (how a visitor sees the site), run without --connect and with --headless, and say so.
• When a script fails, the page stays where it was: reconnect, take a screenshot, log the url and title, then continue from there.

Reply with what you saw or did, the final URL, and the screenshot path when it helps.