# Browser > Use the managed Puffer browser tab through the internal CLI to inspect pages, open tabs, and interact with UI through snapshots and refs. - Skill: `berabuddies/browser` (Agent Skill) - Install (CLI): `npx skillmds@latest add berabuddies/browser` - Raw SKILL.md: https://api.skillmd.com/api/skills/berabuddies/browser/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: berabuddies (https://skillmd.com/u/berabuddies) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/berabuddies/browser --- Use Bash to run the Browser internal CLI when a task requires a real page in Puffer's managed CEF/Chromium browser tab: opening a page, switching tabs, inspecting visible UI, clicking controls, filling fields, pressing keys, uploading files, capturing screenshots, evaluating page JavaScript, or checking page text. Browser is not a model tool and must not be requested as a provider tool call. Run Browser commands as `browser ...` inside Bash. Target: $target Workflow: 1. Resolve login credentials when the task involves signing in. - Before asking the user for credentials or attempting manual login, call `RequestSecret` with `action: "search"` using the site name, domain, origin, visible username/email hint, or login URL. - If exactly one relevant credential is available, request it by id/name and use the returned non-secret `username` metadata for email, phone, username, or account fields. Use the returned `PUFFER_SECRET_...` placeholder only in the browser command that fills the password, token, or other secret-value field. Never print the secret value. - If multiple matches are available, choose only when the metadata is unambiguous; otherwise ask the user which stored credential to use. - If no matching credential exists and login cannot proceed without one, call `RequestSecret` with `action: "collect"`, a clear `name`, `origin`, optional `username`, and a short `prompt` so the user can enter the value into a masked secret prompt. Use the returned `PUFFER_SECRET_...` placeholder; do not ask the user to paste passwords, tokens, cookies, or API keys into chat or `AskUserQuestion`. 2. Open or find the tab. - `browser list` lists tabs for the current agent session. - `browser open https://example.com --label docs --width 1280 --height 900` opens or reuses a managed browser tab. - `browser tab new https://example.com --label scratch` forces a fresh tab. - `browser tab focus t1` switches the active agent-facing tab handle. - Treat DOM readiness as the page-ready signal; a single snapshot is enough to start working. Do not wait for full network idle unless the user explicitly asks to debug network quiescence or a page-specific workflow requires all background requests to stop. 3. Snapshot once to get your bearings. - `browser snapshot --tab-id t1` returns the visible text and fresh refs like `@e1`. - One snapshot lists every interactive ref currently in the viewport — enough to plan a whole form or section. You do not need to snapshot before each field. - Refs stay valid until the page changes. Re-snapshot only after a navigation or when an action reports the page changed — not after every action (see step 5). 4. Act on refs in batches, and trust the result. - Fill or click every field in a form section back-to-back from the same snapshot, then move on — do not snapshot between fields. - Each action returns its result: many actions report the resulting page state (current url, title, and refs); read that returned state to decide your next move instead of taking a separate snapshot. When an action returns only a brief success confirmation, trust it — a silent success means it worked. - `browser click @e3 --tab-id t1` clicks an element from the latest snapshot. - `browser focus @e3 --tab-id t1` focuses an element without clicking it. - `browser fill @e5 "hello" --tab-id t1` replaces text in an editable control. Use this for known final values in email, username, password, search, and address fields. - `browser type "hello" --ref @e5 --tab-id t1` focuses a ref and inserts/appends text. Do not use this to replace a field value. - `browser scroll-into-view @e5 --tab-id t1` centers a ref before interacting when needed. - `browser select @e6 "New York" --tab-id t1` chooses one native `