Browser automation in OpenClaw — the full picture.

OpenClaw's browser is a real Chrome instance the agent can drive. It opens URLs, clicks, types, scrolls, takes screenshots, fills forms, and reads rendered pages — the same way a human would, just deterministically and at agent speed.

Three things make this materially different from a "scrape a webpage" tool:

• JavaScript runs. Single-page apps, lazy loading, modal dialogs — all work because Chrome actually renders the page.
• Sessions persist. The agent can log into a service, then operate inside it across many turns.
• The agent gets a structured page model. Not raw HTML — a snapshot with role-based affordances (button "Submit", link "Next page") so it can decide what to click without parsing CSS selectors.

Out of the box, the gateway spawns a headless Chrome and exposes browser tools to the agent. No setup needed.

> Find the latest Hacker News post about OpenClaw and summarize it.

[agent calls browser_navigate]
url: https://news.ycombinator.com/from?site=openclaw.ai

[agent calls browser_snapshot]
returns: structured listing of stories...

[agent calls browser_navigate]
url: <story URL>

[agent calls browser_get_text]
returns: rendered post content

[agent replies with summary]

Configuration knobs that matter:

{
  "browser": {
    "mode": "managed",
    "headless": true,
    "userAgent": "Mozilla/5.0 (compatible; OpenClawAgent/1.0)",
    "viewport": { "width": 1280, "height": 800 },
    "timeout": 30000,
    "blockResources": ["image", "media"]
  }
}

blockResources is the biggest lever for cost and speed — blocking images cuts page load time 40–60% and shrinks the snapshot tokens significantly.

Added in OpenClaw 2026.3. The agent connects to a Chrome you're already running with remote debugging on. Your logins, extensions, and tabs are all available.

# Start Chrome with remote debugging
google-chrome --remote-debugging-port=9222 --user-data-dir=$HOME/.chrome-openclaw

# Tell OpenClaw to use it
openclaw browser attach ws://localhost:9222

For complex flows where you need precise control — multi-context isolation, advanced waiting, custom event handling — install the Playwright skill from ClawHub.

openclaw skills install playwright

The skill exposes a richer surface: browser_evaluate (run JS in the page), browser_choose_file (file upload), browser_press (key sequences), browser_select_option (dropdowns by value). The trade-off is your agent now has to reason about CSS selectors instead of just role-based affordances — sometimes worth it, often overkill.

The default browser tool returns structured snapshots rather than raw HTML or pixel screenshots. This is the single most important design choice for LLM-driven browsing.

page: example.com/dashboard
heading: "Welcome back, Sam"
nav:
  - link "Projects" → /projects
  - link "Settings" → /settings
  - button "Sign out"
main:
  list "Recent activity":
    - item "Q2 launch · updated 2h ago"
    - item "Hiring · updated yesterday"
  button "New project" → primary
form "Quick add":
  - input "title" (required)
  - select "team" [Eng, PM, Design]
  - button "Add"

The agent can reason "click the 'New project' primary button" instead of "find a button with class .btn-primary inside .dashboard-actions[data-test=ka-new]" — which is fragile and token-heavy. Snapshots are the killer feature.

You will hit them. Treat them as expected.

• Always set a navigation timeout. 30s default is fine; without one a slow page hangs the agent.
• Block images and media when possible. 40–60% faster, much cheaper.
• Use snapshot mode, not raw DOM. Tokens matter; agents reason better on structured data.
• Persist cookies. Make every session ephemeral and you'll re-login constantly. Save browser/cookies.json to the workspace.
• Don't run browser tools at heartbeat speed. Snapshots are token-heavy; reserve browser calls for user-initiated and scheduled work.
• Snapshot before every action. Pages change between turns; stale state causes click-the-wrong-thing errors.
• Use Playwright when selectors are unavoidable. Snapshots can't always disambiguate three buttons that look identical to the role layer.