playwright-mcp-automation（Playwright MCP 自动化）

Playwright MCP Automation

Overview

Use this skill whenever an agent must drive a real browser session via the Playwright MCP server. It covers standing up the MCP daemon, wiring it into your MCP client, and running reliable automation loops (login → navigate → act → verify). Pair these instructions with the upstream repo (<https://github.com/microsoft/playwright-mcp>) for the latest binaries.

Bundled resources

• references/setup.md — launch recipes, flags, troubleshooting.
• references/tools.md — quick lookup table for MCP tools.
• scripts/start_playwright_mcp.sh — opinionated launcher (persistent profile, sane timeouts). Override via env vars or CLI flags.

Quick start (once per host)

1. Install prerequisites

- Ensure Node.js ≥ 18. - Install Playwright browsers and system deps once:

     npx playwright install chromium
     # Linux only: installs missing libraries (x11, fonts, etc.)
     sudo npx playwright install-deps chromium

1. Launch the MCP server

- Fast path: run scripts/start_playwright_mcp.sh from this skill directory. Override with PWMCP_BROWSER, PWMCP_PORT, etc. - Need custom flags? Copy-paste from references/setup.md §2–5.

1. Register the server with your agent

- Local STDIO client:

     {
       "mcpServers": {
         "playwright": {
           "command": "npx",
           "args": ["@playwright/mcp@latest", "--browser=chromium", "--user-data-dir=/home/ai/.cache/playwright/mcp-profile", "--allowed-hosts=*", "--snapshot-mode=incremental"]
         }
       }
     }

- Remote HTTP transport: expose --port/--host then set "url": "http://HOST:PORT/mcp".

1. Sanity check

- Call browser_navigate to https://example.com, then browser_snapshot. If the tree renders, automation is ready.

Core workflow

Follow this loop for every task. Refer to references/tools.md for exact tool signatures.

1. Plan & prep

• Clarify success criteria (e.g., "reach checkout confirmation", "download CSV").
• Decide on headed vs headless mode. Use headed when debugging CAPTCHAs/visual states.
• Make sure secrets (login cookies, OTP hooks) are available. If not, fall back to manual storage-state provisioning.

2. Navigate & observe

1. browser_navigate to the starting URL.
2. Immediately browser_snapshot to capture the accessibility tree.
3. If layout depends on viewport/device, adjust via launch flags (--device, --viewport-size) or call browser_resize.

3. Interact deterministically

Use semantic tools whenever possible:

• Inputs/buttons: browser_click, browser_type, browser_fill_form.
• Dropdowns: browser_select_option.
• Dynamic menus/tooltips: browser_hover → browser_wait_for.
• Multi-step flows (wizards, carts): after every step, browser_snapshot + assert expected text before proceeding.
• When markup lacks good roles, enable --caps=vision and fall back to browser_mouse_* tools, but only as a last resort.
• For brittle sequences (e.g., injecting JS, intercepting fetch), wrap logic inside browser_run_code:

  async (page) => {
    await page.waitForSelector('text=Place order');
    await page.getByRole('button', { name: 'Place order' }).click();
    return await page.getByTestId('order-number').innerText();
  }

4. Handle waits & retries

• Prefer browser_wait_for with text / textGone over arbitrary sleeps.
• On timeout, fetch a fresh snapshot, confirm element existence, then retry once before escalating.
• Capture console/network logs (browser_console_messages, browser_network_requests) to debug API errors or CSP blocks.

5. Verify & capture artefacts

• Collect evidence via browser_snapshot and, if needed, browser_take_screenshot or browser_pdf_save (enable --caps=pdf).
• For multi-tab flows, list tabs via browser_tabs and ensure the correct tab is selected before final actions.
• Always browser_close at the end of unattended runs to release the browser.

Authentication & state strategies

1. Persistent profile (default script)

- Keeps cookies/localStorage inside PWMCP_PROFILE. Great for daily automations. - Rotate profile path per task to avoid cross-site contamination.

1. Storage-state bootstrap

- Use Playwright CLI or manual login to create storage.json. Launch with --storage-state=/path/to/storage.json (see setup reference §4). - Update file whenever passwords change.

1. Secrets file

- Launch with --secrets path/.env so MCP can expose sensitive values via secrets.get. Include API keys or 2FA tokens there instead of SKILL files.

1. Browser extension bridge

- When you must reuse an already-signed-in Chrome profile, install the Playwright MCP Bridge extension and launch with --extension. Follow upstream README for pairing.

Resilience checklist

• CAPTCHA / MFA: Surface snapshots promptly so a human can intervene or provide MFA codes. Document fallback in your agent conversation.
• Slow sites: Increase --timeout-action/--timeout-navigation, or stage requests via browser_wait_for { time }.
• Resource throttling: Use headless mode and disable --save-video/--save-trace unless debugging.
• Logging: Save snapshots to files (browser_snapshot filename) for audit trails, especially when producing evidence (e.g., order confirmations).

When to read the references/scripts

• Need launch flags, network rules, or troubleshooting? → Open references/setup.md.
• Need to recall exact tool names/parameters? → Check references/tools.md.
• Want a ready-to-run launcher? → Execute or adapt scripts/start_playwright_mcp.sh. Export env vars (PWMCP_PORT, PWMCP_EXTRA="--headless") before running to tweak behavior.

Keep SKILL.md lean by offloading details to the references. Update references/scripts whenever the upstream Playwright MCP release adds new capabilities (vision, pdf, devtools, etc.).