Browser Control (Claude Code)

Use bundled Playwright scripts for deterministic web automation.

Prerequisites

• Python 3.10+
• pip install playwright
• python -m playwright install chromium

Scripts

• .claude/skills/browser-control/scripts/open_url.py
• .claude/skills/browser-control/scripts/snapshot_dom.py
• .claude/skills/browser-control/scripts/act_click_type.py
• .claude/skills/browser-control/scripts/wait_for.py
• .claude/skills/browser-control/scripts/screenshot.py

Recommended flow

1. Open target URL
2. Snapshot DOM candidates
3. Run click/type/press actions
4. Wait for selector/text state
5. Capture screenshot and report

Commands

Open URL

python3 .claude/skills/browser-control/scripts/open_url.py \
  --url "https://example.com" \
  --out outputs/browser/open.json

Snapshot candidates

python3 .claude/skills/browser-control/scripts/snapshot_dom.py \
  --url "https://example.com" \
  --out outputs/browser/snapshot.json

Actions (click/type/press)

python3 .claude/skills/browser-control/scripts/act_click_type.py \
  --url "https://example.com" \
  --actions '[{"type":"click","selector":"text=More information"}]' \
  --out outputs/browser/act.json

Wait for UI state

python3 .claude/skills/browser-control/scripts/wait_for.py \
  --url "https://example.com" \
  --selector "h1" \
  --timeout-ms 10000 \
  --out outputs/browser/wait.json

Screenshot

python3 .claude/skills/browser-control/scripts/screenshot.py \
  --url "https://example.com" \
  --image outputs/browser/page.png \
  --out outputs/browser/screenshot.json

Token optimization tips

Each script call and its output consumes tokens. Minimize unnecessary steps:

1. Skip DOM snapshot when URL pattern is known

   • Search engines (Naver, Google) have predictable URL patterns. Build the URL directly instead of snapshotting + typing + pressing Enter.
   bashCopy

   # BAD: 3 steps (snapshot → type/click/press → screenshot) = high token cost
   python3 .claude/skills/browser-control/scripts/snapshot_dom.py --url "https://www.naver.com" ...
   python3 .claude/skills/browser-control/scripts/act_click_type.py --url "https://www.naver.com" --actions '[...]' ...
   python3 .claude/skills/browser-control/scripts/screenshot.py --url "https://search.naver.com/..." ...
   
   # GOOD: 1 step = low token cost
   python3 .claude/skills/browser-control/scripts/screenshot.py \
     --url "https://search.naver.com/search.naver?query=%EC%98%A4%ED%94%88%ED%81%B4%EB%A1%9C" \
     --image outputs/browser/result.png \
     --out outputs/browser/screenshot.json
   

2. Never Read the full snapshot JSON into context

   • Snapshot JSON can be hundreds of lines. Filter with jq in Bash instead of using the Read tool.
   bashCopy

   python3 .claude/skills/browser-control/scripts/snapshot_dom.py \
     --url "https://example.com" --out /dev/stdout \
     | jq '.items[] | select(.tag=="input")'
   

3. Combine actions into a single call

   • Use one act_click_type.py call with multiple actions array instead of calling the script multiple times.

4. Screenshot only when needed

   • If you only need text data (e.g. link URLs, page title), use snapshot_dom.py with jq filtering instead of a screenshot.

Policy (read/write access control)

browser-control/policy.yaml controls domain access.

• write_allowed_hosts: only these hosts/subdomains are writable (deny by default for write)
• read_blocked_hosts: only these hosts/subdomains are blocked for read (all others readable)
• Conflict rule: deny wins
• Redirect rule: policy is checked for both requested URL and final redirected URL

Example policy.yaml (JSON-compatible YAML):

{
  "write_allowed_hosts": ["docs.openclaw.ai", "github.com"],
  "read_blocked_hosts": ["facebook.com", "x.com"]
}

Rules

• Prefer stable selectors (data-testid, explicit ids, role/text) over brittle nth-child selectors.
• Save every step output JSON for reproducibility.
• If an action fails, return the failing selector/action index clearly.
• Use --headed only for debugging; default headless for automation.