Ms Playwright MCP

Playwright MCP Skill

Workflow guidance for Microsoft Playwright MCP server. Tool parameters are delivered by MCP tools/list handshake - this skill provides procedures, gotchas, and decision logic.

References (loaded on demand):

• PLAYWRIGHT_TOOLS.md - Complete tool catalog (21 core + 40 opt-in)
• PLAYWRIGHT_CONFIG.md - All CLI flags, env vars, JSON config
• PLAYWRIGHT_AUTHENTICATION.md - Extension mode, persistent profiles
• PLAYWRIGHT_TROUBLESHOOTING.md - Common issues with fixes
• SETUP.md - Installation
• UNINSTALL.md - Uninstallation

MUST-NOT-FORGET

1. Use browser_snapshot (accessibility tree) for element discovery, not screenshots
2. Element refs are ephemeral - ALWAYS re-snapshot before clicking after any navigation or page change
3. Reference elements via ref from snapshot (e.g., ref: "e5") or selector (e.g., [data-testid="submit"])
4. Use type: "jpeg" for all screenshots (default PNG produces unnecessarily large files)
5. NEVER auto-close browser - only user may close. Sessions preserve authentication state
6. browser_fill_form takes fields array (NOT browser_fill). browser_take_screenshot (NOT browser_screenshot)
7. Opt-in tools need --caps flag. Cookie read needs --caps=storage. Route mocking needs --caps=network
8. browser_evaluate runs in browser (no Node.js APIs). browser_run_code runs server-side with Playwright page object
9. Downloads go to --output-dir path. Default: .playwright-mcp/ under the IDE installation folder (e.g., [LOCALAPPDATA]\Programs\Devin\.playwright-mcp\). NOT the system Downloads folder. File names are typically UUIDs (e.g., 765984ac-5fcb-4bad-9aa3-e4163a93fa5d) or descriptive names from the server. Use Get-ChildItem "[LOCALAPPDATA]\Programs\Devin\.playwright-mcp" -File | Sort-Object LastWriteTime -Descending | Select-Object -First 10 FullName, Length, LastWriteTime to find recent downloads. For PDFs, extract text with poppler: & "[TOOLS]\poppler\Library\bin\pdftotext.exe" "<path>" -
10. Extension mode: Chrome/Edge only, uses Chrome Web Store extension, NOT --remote-debugging-port
11. fullPage: true + type: "jpeg" is MANDATORY for research, archival, or source-validation screenshots. Viewport-only (fullPage: false) captures ~900px of a 5,000-20,000px page - missing 80-95% of content. PNG produces 3-5x larger files than JPEG with no benefit for text pages. Only use viewport for UI debugging or above-the-fold checks. (GLOB-FL-0042)

Intent Lookup

User wants to...

• Read a webpage / research a topic → Navigate, snapshot, dismiss cookie popup, scroll for lazy content
• Fill out a form → Snapshot to get refs, browser_fill_form with fields array, click submit
• Take a screenshot → browser_take_screenshot(type: "jpeg"), or fullPage: true for complete page
• Log into a site → Fill credentials + submit, or use persistent profile. See PLAYWRIGHT_AUTHENTICATION.md
• Use an already logged-in browser → Extension mode (--extension). See PLAYWRIGHT_AUTHENTICATION.md
• Scrape data → browser_evaluate with JS to extract DOM content, or browser_snapshot for structured text
• Test a web UI → Navigate, snapshot, assert elements present, use --caps=testing for verification tools
• Download a file → Find link with snapshot, click to download, check output dir
• Interact with a map / canvas / custom widget → Need --caps=vision for coordinate-based click/drag
• Mock API responses → Need --caps=network, use browser_route with URL pattern
• Handle cookie/General Data Protection Regulation (GDPR) popups → Snapshot, find accept/reject button, click it. See Dismiss Cookie Popup
• Debug browser issues → See PLAYWRIGHT_TROUBLESHOOTING.md

Core Procedures

1. Navigate and Interact

1. browser_navigate(url: "https://example.com")
2. browser_snapshot()                              # Get element refs
3. # If cookie popup visible: find accept button, click it, re-snapshot
4. browser_click(ref: "e12", element: "Target button")
5. browser_snapshot()                              # Verify result, get new refs

After EVERY navigation or click that changes the page: re-snapshot. Refs from previous snapshot are invalid.

2. Fill a Form

1. browser_snapshot()                              # Find form fields
2. browser_fill_form(fields: [
     {ref: "e3", name: "Email", type: "textbox", value: "user@example.com"},
     {ref: "e5", name: "Password", type: "textbox", value: "password123"}
   ])
3. browser_click(ref: "e8", element: "Submit button")
4. browser_snapshot()                              # Verify submission result

Field types: textbox, checkbox (value: "true"/"false"), radio, combobox (value: option text), slider.

3. Full Page Screenshot

1. browser_navigate(url: "https://example.com")
2. browser_snapshot()                              # Check for cookie popups
3. # Dismiss cookie popup if present
4. # Scroll down to trigger lazy-loaded content:
   browser_press_key(key: "End")
   browser_wait_for(time: 2)
   browser_press_key(key: "Home")
5. browser_take_screenshot(fullPage: true, type: "jpeg")

4. Extract Data with JavaScript

1. browser_navigate(url: "https://example.com")
2. browser_evaluate(function: "document.querySelectorAll('h2').length")
3. browser_evaluate(function: "JSON.stringify([...document.querySelectorAll('tr')].map(r => r.textContent))")

browser_evaluate runs in browser context. No Node.js APIs. Return values must be serializable. For Playwright API access (e.g., page.locator()), use browser_run_code instead.

5. Dismiss Cookie/GDPR Popup

1. browser_snapshot()
2. # Look for buttons with text: Accept, Agree, OK, Got it, Allow, Consent
3. browser_click(ref: "<popup-button-ref>", element: "Accept cookies")
4. browser_snapshot()                              # Verify popup dismissed

If popup uses iframe: snapshot may show it nested. Try clicking by text selector: selector: "text=Accept".

6. Scroll and Load Lazy Content

1. browser_press_key(key: "End")                   # Scroll to bottom
2. browser_wait_for(time: 2)                        # Wait for lazy load
3. browser_snapshot()                              # Check if new content appeared
4. # Repeat if more content expected

For infinite scroll: repeat scroll+wait+snapshot in a loop until content stops changing.

Element Targeting

Two methods (v0.0.69+):

1. ref from snapshot (preferred) - ref: "e5" from latest browser_snapshot
2. selector (when refs unavailable) - priority order:
   • [data-testid="submit"] - Test IDs (most stable)
   • role=button[name="Save"] - Semantic roles
   • text=Sign in - Visible text
   • input[name="email"] - HTML attributes

All interaction tools (browser_click, browser_type, browser_hover, etc.) accept both ref and selector.

Capabilities Quick Reference

Core 21 tools always available. Opt-in categories enabled via --caps:

• network - Route mocking, offline toggle
• storage - Cookie/localStorage/sessionStorage Create/Read/Update/Delete (CRUD)
• devtools - Tracing, video recording, debugger
• vision - Coordinate-based mouse (for canvas/maps)
• pdf - Save page as PDF
• testing - Element/text/value verification, locator generation
• config - Read resolved server configuration

Enable all: --caps vision,pdf,devtools,network,storage,testing,config. Tool details: PLAYWRIGHT_TOOLS.md

Gotchas

• Refs invalidated by any page change - navigation, DOM mutation, dynamic content. Always re-snapshot before interacting
• browser_fill_form not browser_fill - Old name deprecated. New API takes fields array, not single field
• browser_take_screenshot not browser_screenshot - Old name deprecated
• browser_evaluate runs in BROWSER - No Node.js. Use browser_run_code for server-side Playwright API
• Cookie read needs --caps=storage - Even basic browser_cookie_list is gated
• browser_network_requests is core - Read-only request listing. Route mocking needs --caps=network
• Extension mode is Chrome/Edge only - Firefox/WebKit not supported
• --no-sandbox warning is cosmetic - Browser works correctly. Required in Docker/WSL2
• Default screenshot format is PNG - Always specify type: "jpeg" for smaller files
• Profile lock error - Previous Chrome didn't shut down cleanly. Close Chrome instances or delete lock file

Quick Config

Basic (persistent profile, default):

{"mcpServers": {"playwright": {"command": "npx", "args": ["@playwright/mcp@latest"]}}}

Persistent profile with all capabilities:

{"mcpServers": {"playwright": {"command": "npx", "args": ["@playwright/mcp@latest", "--caps", "vision,pdf,devtools,network,storage,testing,config"]}}}

Extension mode (existing browser):

{"mcpServers": {"playwright": {"command": "npx", "args": ["@playwright/mcp@latest", "--extension"]}}}

Full config reference: PLAYWRIGHT_CONFIG.md