vipin wiki

Chrome Automation Skill

3 min read

Chrome Automation Skill

Role In The Wiki

chrome-automation is a project-local installed skill for controlling real Google Chrome through agent-browser and Chrome DevTools Protocol.

Current Claims

  • EXTRACTED: The upstream path is https://github.com/anbeime/skill/tree/main/skills/chrome-automation/chrome-automation.
  • EXTRACTED: It was mirrored under skill/chrome-automation/.
  • EXTRACTED: It was installed under .codex/skills/chrome-automation/.
  • EXTRACTED: The installed skill was locally annotated with a D-drive runtime override because upstream examples assume ~/.claude/skills and $HOME\Documents\agent-browser.
  • EXTRACTED: The D-drive agent-browser-win32-x64.exe runtime passed a real CDP smoke test on 2026-05-17.

Usage

Use it when the user needs visible real-browser automation:

用 chrome-automation 打开这个网站并截图。
Use chrome-automation to fill this form in real Chrome.
控制我当前 Chrome 页面，帮我点这个按钮。

Local command pattern:

$AB_HOME = "D:\Research\vipin's knowledgebase\.wiki-tmp\tools\agent-browser"
$AB_BIN = "$AB_HOME\bin\agent-browser-win32-x64.exe"
$PORT = 9223
$env:AGENT_BROWSER_HOME = $AB_HOME
& $AB_BIN open https://example.com --cdp $PORT --session codex-smoke
& $AB_BIN get title --cdp $PORT --session codex-smoke
& $AB_BIN get url --cdp $PORT --session codex-smoke
& $AB_BIN snapshot -i --cdp $PORT --session codex-smoke
& $AB_BIN screenshot body ".wiki-tmp\state.png" --cdp $PORT --session codex-smoke

Before use, ensure Chrome is running with --remote-debugging-port=$PORT.

Verification Record

  • EXTRACTED: Installed agent-browser under .wiki-tmp/tools/agent-browser/.
  • EXTRACTED: Downloaded binary returned agent-browser 0.27.0.
  • EXTRACTED: Chrome launched with --remote-debugging-port=9223 and D-drive profile .wiki-tmp/tools/chrome-automation-profile.
  • EXTRACTED: agent-browser --cdp 9223 open https://example.com succeeded.
  • EXTRACTED: get title returned Example Domain.
  • EXTRACTED: get url returned https://example.com/.
  • EXTRACTED: screenshot wrote .wiki-tmp/chrome-automation-example.png.
  • EXTRACTED: A retest found the Windows binary is most reliable when global flags are placed after the command, for example open https://example.com --cdp 9224 --session codex-retest.
  • EXTRACTED: Saved screenshots work with screenshot body <path> or screenshot --full <path>; screenshot <path> can be parsed as a selector and fail.
  • EXTRACTED: On 2026-05-17, the D-drive agent-browser runtime completed a live Feishu shared Base form workflow through feishu-bridge: filled fields, handled stubborn select controls, passed required-field validation, clicked the one-time-submit confirmation, and verified final 提交成功 page text.
  • EXTRACTED: For this Feishu form, snapshot -i, DOM inspection, coordinate clicking, and Escape to close open dropdowns were all needed; a simple “click visible option” flow was not reliable enough.
  • AMBIGUOUS: Native Rust rebuild failed because MSVC link.exe was missing; this does not block the downloaded Windows binary.

Safety Boundary

  • Do not enter credentials or publish/submit forms without explicit user confirmation.
  • Prefer isolated D-drive Chrome profiles for testing.
  • Use real-user sessions only when the user asks for that context.
  • Show screenshots or summarize visible state after significant browser actions.

Counterpoints And Gaps

  • AMBIGUOUS: Upstream auto-install scripts still default to $HOME\Documents; future agents should use the D-drive override unless the user explicitly wants global installation.
  • AMBIGUOUS: The upstream docs often show --cdp <port> before the command; on this local Windows runtime, use command-first examples in this page.
  • INFERRED: If a future task requires native rebuild, installing Visual Studio Build Tools with MSVC linker may be necessary; this should not be committed to the repo.

Graph View

Backlinks