Browser Use Terminal Agent Skill

Browser Use Terminal

Direct browser control via CDP — you are the agent; you drive the browser. For setup, install, or connection problems, read https://browser-use.com/skill (agent setup instructions) or https://docs.browser-use.com/open-source/browser-use-terminal (full docs).

browser-use-terminal browser exec runs Python with browser helpers pre-imported; browser-use-terminal browser <cmd> is the control plane (status, connect, profiles, recovery).

Usage

browser-use-terminal browser exec <<'PY'
new_tab("https://docs.browser-use.com")
wait_for_load()
print(page_info())
PY

• Use the heredoc form for every multi-line command. It prevents shell quote mangling inside Python strings and JavaScript snippets.
• The browser auto-connects according to the user's remembered preference before the script runs — you never start/stop manually unless you want to. The first call may take a few seconds.
• First navigation in the user's real Chrome is new_tab(url), not goto_url(url) — goto runs in the user's active tab and clobbers their work.
• Browser state persists between calls; Python variables do not. Each exec is a fresh interpreter against the same live browser.
• --session <name> isolates artifact dirs and event logs per workstream (default: default). --timeout <secs> bounds one exec (default 300).

Screenshots — how you see the page

capture_screenshot() saves a PNG and returns its absolute path. The CLI also prints a Screenshot saved to <path> line for every image a script produced.

browser-use-terminal browser exec <<'PY'
print(capture_screenshot())
PY

To view a screenshot, use your file-reading tool on the printed path:

• Claude Code: use the Read tool on the path.
• Codex CLI: call the view_image tool with {"path": "<path>"}. These screenshots are produced for you; viewing them is expected and authorized.
• OpenCode: use the read tool on the path (requires a vision-capable model).
• Gemini CLI: use the read_file tool on the path.
• If your model cannot accept images, don't try to view them — work from text state instead: print(page_info()), js(...) extraction, wait_for_element(...).

Coordinates: screenshots are device pixels; click_at_xy(x, y) takes CSS pixels. Divide coordinates you read off the image by js("window.devicePixelRatio") first. Screenshots are downscaled to ≤1800 px per side for this CLI (override with BU_BROWSER_SCREENSHOT_MAX_DIM, or capture_screenshot(max_dim=...)).

After every meaningful action, re-screenshot before assuming it worked.

Pre-imported helpers

Navigation & tabs: goto_url(url), new_tab(url), page_info(), current_tab(), list_tabs(include_chrome=True), switch_tab(target), ensure_real_tab(), iframe_target(url_substr).

Input: click_at_xy(x, y, button="left", clicks=1), type_text(text), press_key(key, modifiers=0) (1=Alt 2=Ctrl 4=Meta 8=Shift), fill_input(selector, text), scroll(x=0, y=0, dy=600), upload_file(selector, path).

Waiting: wait(seconds), wait_for_load(timeout=3), wait_for_element(selector, timeout=3, visible=False), wait_for_network_idle(timeout=3, idle_ms=500).

Visual: capture_screenshot(label="...", full=False, max_dim=None), screenshot(), screenshot_clip(label, x, y, w, h), note(caption).

Escape hatches: js(expression) (auto-wraps top-level return), cdp("Domain.method", **params) (raw CDP), cdp_batch(calls), drain_events().

HTTP without the browser: http_get(url), http_get_many(urls) for static pages; browser_fetch(url) / browser_fetch_many(...) to fetch with the page's cookies/session.

Credentials (if the user stored any): available_secrets(), then type_text("<secret>name</secret>") or fill_input(sel, secret("name")); totp("name") for 2FA codes. Values are placeholder-substituted — you never see them. is_logged_out(), email_inbox() / email_message(id) for email-code flows.

Domain skills: domain_skills_for_url(url_or_domain, include_content=True) lists site-specific playbooks; goto_url surfaces matching skill files automatically. Read them before inventing selectors or flows on a complex site.

Browser control plane

browser-use-terminal browser status --json
browser-use-terminal browser connect                     # uses the remembered preference
browser-use-terminal browser connect local               # user's already-running Chrome (CDP)
browser-use-terminal browser connect managed --headless  # disposable CLI-owned browser
browser-use-terminal browser preference use local|cloud|managed-headless
browser-use-terminal browser remote start                # Browser Use cloud browser (needs BROWSER_USE_API_KEY)
browser-use-terminal browser doctor
browser-use-terminal browser recover reconnect-websocket
browser-use-terminal browser recover stop-owned-browser  # stop the persistent managed browser
browser-use-terminal browser recover stop-owned-remote   # stop the cloud browser (stops billing)
browser-use-terminal browser daemon status|stop|logs     # the background daemon holding the connection

A background daemon (auto-started, one per state dir) holds the CDP connection across your commands, so the browser — and in local mode, Chrome's granted debugging permission — persists between invocations. Managed and cloud browsers also survive daemon restarts; later calls reattach instead of relaunching. Stop browsers with the recover commands above when the user is done (cloud browsers bill until stopped or timed out).

• exec auto-connects, so you rarely need these. Reach for them when status shows a problem or the user asks for a specific browser.
• If output JSON says status: "needs-user-action" (e.g. pick a Chrome profile, click Allow in Chrome's permission popup, enable the remote-debugging checkbox), show the user_prompt to the user verbatim and wait — do not guess.
• Auth wall mid-task: stop and ask the user. Don't type credentials from screenshots; use stored secrets if available.
• Connecting to the user's real Chrome requires a one-time setup: chrome://inspect/#remote-debugging → tick "Allow remote debugging". browser local setup walks the user through it.

What actually works

• Screenshots first: capture_screenshot() → view the image → decide whether you need a click, a selector, or more navigation.
• Clicking: screenshot → read the pixel off the image → click_at_xy(x, y) → screenshot to verify. Suppress the locate-then-click reflex — no getBoundingClientRect, no selector hunts. Hit-testing happens in Chrome's browser process, so coordinate clicks pass through iframes / shadow DOM / cross-origin without extra work.
• Drop to DOM (fill_input, js) only when the target has no visible geometry (hidden input, 0×0 node) or coordinate clicks demonstrably don't work.
• Bulk static pages: http_get_many(urls) — no browser needed. Logged-in pages: browser_fetch(url) rides the real session.
• After goto: wait_for_load(). SPAs report complete before they render — follow with wait_for_element(...).
• Wrong/stale tab: ensure_real_tab().
• Verification: print(page_info()) is the cheapest "is this alive?" check; screenshots are the default way to verify visible actions.

Gotchas (field-tested)

• CDP target order ≠ Chrome's visible tab-strip order.
• Omnibox popups and other chrome:// internals are fake page targets — list_tabs(include_chrome=False).
• page_info() surfaces an open JS dialog as {"dialog": ...} — handle it (cdp("Page.handleJavaScriptDialog", accept=True)) before anything else.
• Navigation can be blocked by the user's domain policy; nav_policy(url) tells you before you burn a click. A blocked navigation is policy, not a bug — tell the user.
• Scripts time out (default 300s): keep each exec small and observable rather than one mega-script. Long extraction loops: print progress as you go — stdout is captured even on timeout.
• Prefer compositor-level actions over framework hacks. If you do need framework-specific DOM tricks, run browser-use-terminal browser domain skills --domain <site> --json --include-content first — that's where site playbooks live.