name: browser description: Direct browser control via CDP. Use when the user wants to automate, scrape, test, or interact with web pages. Connects to the user's already-running Chrome.

browser-harness

Direct browser control via CDP. For task-specific edits, use agent-workspace/agent_helpers.py. For setup, install, or connection problems, read install.md.

Domain skills (community-contributed per-site playbooks under agent-workspace/domain-skills/) are off by default. Set BH_DOMAIN_SKILLS=1 to enable them; see the bottom section.

If BH_DOMAIN_SKILLS=1 and the task is site-specific, read every file in the matching agent-workspace/domain-skills/<site>/ directory before inventing an approach.

Usage

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

• Invoke as browser-harness — it's on $PATH. No cd, no uv run.
• Use the heredoc form for every multi-line command. It prevents shell quote mangling inside Python strings and JavaScript snippets.
• First navigation is new_tab(url), not goto_url(url) — goto runs in the user's active tab and clobbers their work.

Tool call shape

browser-harness <<'PY'
# any python. helpers pre-imported. daemon auto-starts.
PY

run.py calls ensure_daemon() before exec — you never start/stop manually unless you want to.

Remote browsers

Use remote for parallel sub-agents (each gets its own isolated browser via a distinct BU_NAME) or on a headless server. BROWSER_USE_API_KEY must be set. start_remote_daemon, list_cloud_profiles, list_local_profiles, sync_local_profile are pre-imported.

browser-harness <<'PY'
start_remote_daemon("work")                               # default — clean browser, no profile
# start_remote_daemon("work", profileName="my-work")      # reuse a cloud profile (already logged in)
# start_remote_daemon("work", profileId="<uuid>")         # same, but by UUID
# start_remote_daemon("work", proxyCountryCode="de", timeout=120)   # DE proxy, 2-hour timeout
# start_remote_daemon("work", proxyCountryCode=None)      # disable the Browser Use proxy
PY

BU_NAME=work browser-harness <<'PY'
new_tab("https://example.com")
print(page_info())
PY

start_remote_daemon prints liveUrl and auto-opens it in the local browser (if a GUI is detected) so the user can watch along. Headless servers print only — share the URL with the user. The daemon PATCHes the cloud browser to stop on shutdown, which persists profile state. Running remote daemons bill until timeout.

Profiles (cookies-only login state) live in interaction-skills/profile-sync.md — covers list_cloud_profiles(), the chat-driven "which profile?" pattern, and sync_local_profile() for uploading a local Chrome profile.

Interaction skills

If you start struggling with a specific mechanic while navigating, look in interaction-skills/ for helpers. They cover reusable UI mechanics like dialogs, tabs, dropdowns, iframes, and uploads. The available interaction skills are:

• connection.md
• cookies.md
• cross-origin-iframes.md
• dialogs.md
• downloads.md
• drag-and-drop.md
• dropdowns.md
• iframes.md
• network-requests.md
• print-as-pdf.md
• profile-sync.md
• screenshots.md
• scrolling.md
• shadow-dom.md
• tabs.md
• uploads.md
• viewport.md

What actually works

• Screenshots first: use capture_screenshot() to understand the current page quickly, find visible targets, and decide whether you need a click, a selector, or more navigation.
• Clicking: capture_screenshot() → read the pixel off the image → click_at_xy(x, y) → capture_screenshot() to verify. Suppress the Playwright-habit reflex of "locate first, then click" — no getBoundingClientRect, no selector hunt. Drop to DOM only when the target has no visible geometry (hidden input, 0×0 node). Hit-testing happens in Chrome's browser process, so clicks go through iframes / shadow DOM / cross-origin without extra work.
• Bulk HTTP: http_get(url) + ThreadPoolExecutor. No browser for static pages (249 Netflix pages in 2.8s).
• After goto: wait_for_load().
• Wrong/stale tab: ensure_real_tab(). Use it when the current tab is stale or internal; the daemon also auto-recovers from stale sessions on the next call.
• Verification: print(page_info()) is the simplest "is this alive?" check, but screenshots are the default way to verify whether a visible action actually worked.
• DOM reads: use js(...) for inspection and extraction when the screenshot shows that coordinates are the wrong tool.
• Iframe sites (Azure blades, Salesforce): click_at_xy(x, y) passes through; only drop to iframe DOM work when coordinate clicks are the wrong tool.
• Auth wall: redirected to login → stop and ask the user. Don't type credentials from screenshots.
• Raw CDP for anything helpers don't cover: cdp("Domain.method", params).

Design constraints

• Coordinate clicks default. Input.dispatchMouseEvent goes through iframes/shadow/cross-origin at the compositor level.
• Connect to the user's running Chrome. Don't launch your own browser.
• cdp-use is only for CDPClient.send_raw. Prefer raw CDP strings over typed wrappers.
• run.py stays tiny. No argparse, subcommands, or extra control layer.
• Core helpers stay short. Put task-specific helper additions in agent-workspace/agent_helpers.py; daemon/bootstrap and remote session admin live in the core package.
• Don't add a manager layer. No retries framework, session manager, daemon supervisor, config system, or logging framework.

Gotchas (field-tested)

• Omnibox popups are fake page targets. Filter chrome://omnibox-popup... and other internals when you need a real tab.
• CDP target order != Chrome's visible tab-strip order. Use UI automation when the user means "the first/second tab I can see"; Target.activateTarget only shows a known target.
• Default daemon sessions can go stale. ensure_real_tab() re-attaches to a real page.
• Browser Use API is camelCase on the wire. cdpUrl, proxyCountryCode, etc.
• Remote cdpUrl is HTTPS, not ws. Resolve the websocket URL via /json/version.
• Stop cloud browsers with PATCH /browsers/{id} + {"action":"stop"}.
• After every meaningful action, re-screenshot before assuming it worked. Use the image to verify changed state, open menus, navigation, visible errors, and whether the page is in the state you expected.
• Use screenshots to drive exploration. They are often the fastest way to find the next click target, notice hidden blockers, and decide if a selector is even worth writing.
• Prefer compositor-level actions over framework hacks. Try screenshots, coordinate clicks, and raw key input before adding DOM-specific workarounds.
• If you need framework-specific DOM tricks, check interaction-skills/ first. That is where dropdown, dialog, iframe, shadow DOM, and form-specific guidance belongs.

Domain skills (opt-in)

Only applies when BH_DOMAIN_SKILLS=1. Otherwise ignore — agent-workspace/domain-skills/ is dormant and goto_url won't surface skill files.

When enabled, search agent-workspace/domain-skills/<host>/ before inventing an approach. goto_url returns up to 10 skill filenames for the navigated host.

If you learn anything non-obvious — a private API, stable selector, framework quirk, URL pattern, hidden wait, or site-specific trap — open a PR to agent-workspace/domain-skills/<site>/. Capture the durable shape of the site (the map, not the diary). Don't write pixel coordinates (break on layout), task narration, or secrets — the directory is public.

Browser Harness ♞

Connect an LLM directly to your real browser with a thin, editable CDP harness. For browser tasks where you need complete freedom.

One websocket to Chrome, nothing between. The agent writes what's missing during execution. The harness improves itself every run.

  ● agent: wants to upload a file
  │
  ● agent-workspace/agent_helpers.py → helper missing
  │
  ● agent writes it                         agent_helpers.py
  │                                                       + custom helper
  ✓ file uploaded

You will never use the browser again.

Setup prompt

Paste into Claude Code or Codex:

Set up https://github.com/browser-use/browser-harness for me.

Read `install.md` and follow the steps to install browser-harness and connect it to my browser.

The agent will open chrome://inspect/#remote-debugging. Tick the checkbox so the agent can connect to your browser:

Click Allow when the per-attach popup appears (Chrome 144+):

See agent-workspace/domain-skills/ for example tasks.

Free Browser Use Cloud browsers

Stealth, sub-agents, or headless deployment. Browser Use Cloud free tier: 3 concurrent browsers, proxies, captcha solving, and more. No card required.

• Grab a key at cloud.browser-use.com/new-api-key
• Or let the agent sign up itself via docs.browser-use.com/llms.txt (setup flow + challenge context included).

Want the deployed version: a 24/7 Linux box agent with Telegram control and a persistent cloud browser? See Browser Use Box and watch the demo.

Architecture (~1k lines across 4 core files)

• install.md — first-time install and browser bootstrap
• SKILL.md — day-to-day usage
• src/browser_harness/ — protected core package
• agent-workspace/agent_helpers.py — helper code the agent edits
• agent-workspace/domain-skills/ — reusable site-specific skills the agent edits

Contributing

PRs and improvements welcome. The best way to help: contribute a new domain skill under agent-workspace/domain-skills/ for a site or task you use often (LinkedIn outreach, ordering on Amazon, filing expenses, etc.). Each skill teaches the agent the selectors, flows, and edge cases it would otherwise have to rediscover.

• Skills are written by the harness, not by you. Just run your task with the agent — when it figures something non-obvious out, it files the skill itself (see SKILL.md). Please don't hand-author skill files; agent-generated ones reflect what actually works in the browser.
• Open a PR with the generated agent-workspace/domain-skills/<site>/ folder — small and focused is great.
• Bug fixes, docs tweaks, and helper improvements are equally welcome.
• Browse existing skills (github/, linkedin/, amazon/, ...) to see the shape.

If you're not sure where to start, open an issue and we'll point you somewhere useful.

Domain skills

Set BH_DOMAIN_SKILLS=1 to enable agent-workspace/domain-skills/ — community-contributed per-site playbooks goto_url surfaces by domain. Contribute via PR.

The Bitter Lesson of Agent Harnesses · Web Agents That Actually Learn