Chrome devtools

You are helping the user interact with websites using Chrome via the chrome-devtools MCP servers.

There are 4 isolated server instances available. You must pick a free one before doing any browser work.

Progress Checklist

Copy this checklist to track your progress through the workflow:

Chrome DevTools Progress:
- [ ] Step 1: Selected an available server instance
- [ ] Step 2: Loaded required MCP tools via ToolSearch
- [ ] Step 3: Performed the requested browser task
- [ ] Step 4: Cleaned up (closed pages)

Step 1: Select an Available Server

There are 4 MCP server instances — 3 headless and 1 visible (opens a real browser window):

Instance	Tool prefix	Mode
1	mcp__chrome-devtools__*	Headless
2	mcp__chrome-devtools-2__*	Headless
3	mcp__chrome-devtools-3__*	Headless
visible	mcp__chrome-devtools-visible__*	Visible

When to use visible: Use chrome-devtools-visible when the user asks to see the browser, needs to debug visually, or wants to watch automation in real-time. Default to headless instances for everything else.

To find a free instance, check each server's list_pages tool (e.g. mcp__chrome-devtools__list_pages). Load them via ToolSearch:

ToolSearch: "+chrome-devtools list_pages"

Then call list_pages on each instance. A server is free if it only has the default about:blank page (or returns no pages). A server is busy if it has pages with real URLs.

Check headless servers in order (1 → 2 → 3) and use the first free one. Only use visible if the user explicitly asks for it or all headless instances are busy.

If all 4 are busy, tell the user:

"All chrome-devtools instances are currently in use. Wait for another task to finish or add more instances."

CRITICAL: Never run parallel Task agents on the same server instance. Each agent must claim its own instance.

Step 2: Load MCP Tools

Once you've picked a server (e.g. instance 2), load the tools you need via ToolSearch using the correct prefix:

ToolSearch: "+chrome-devtools-2 navigate"
ToolSearch: "+chrome-devtools-2 screenshot"

Common tools available on each instance (shown for instance 1, substitute prefix for others):

• mcp__chrome-devtools__navigate_page — Navigate to a URL
• mcp__chrome-devtools__take_screenshot — Capture the current page
• mcp__chrome-devtools__take_snapshot — Get the DOM/accessibility snapshot (useful for scraping text content)
• mcp__chrome-devtools__click — Click an element
• mcp__chrome-devtools__fill — Fill a form field
• mcp__chrome-devtools__fill_form — Fill multiple form fields
• mcp__chrome-devtools__evaluate_script — Run JavaScript on the page
• mcp__chrome-devtools__press_key — Press a keyboard key
• mcp__chrome-devtools__wait_for — Wait for an element or condition
• mcp__chrome-devtools__new_page — Open a new page/tab
• mcp__chrome-devtools__close_page — Close a page/tab
• mcp__chrome-devtools__list_pages — List open pages
• mcp__chrome-devtools__list_network_requests — Monitor network traffic
• mcp__chrome-devtools__list_console_messages — Read console output

Step 3: Perform the Browser Task

Navigation

Always start by navigating to the target URL:

navigate_page(url: "https://example.com")

After navigation, take a screenshot or snapshot to verify the page loaded correctly.

Taking Screenshots

Use take_screenshot to capture what the page looks like. This is useful for:

• Verifying a page loaded correctly
• Debugging layout issues
• Showing the user what you see

Scraping Content

Use take_snapshot to get a structured text representation of the page. This gives you the accessibility tree which is more useful than raw HTML for extracting content.

For dynamic content, you may need to:

1. wait_for the content to appear
2. Then take_snapshot

Form Interaction

1. Use take_snapshot to identify form fields
2. Use fill or fill_form to enter data
3. Use click to submit

Running JavaScript

Use evaluate_script for operations that don't have dedicated tools:

evaluate_script(expression: "document.title")

Step 4: Clean Up

Always close pages when done. There is no auto-timeout on these servers.

close_page(pageId: "<id>")

If you opened multiple pages, close all of them. Use list_pages to verify nothing is left behind.

Troubleshooting

Server won't respond / tools not loading

The MCP server may not be running. Try:

1. Check if the server is listed: claude mcp list --scope user
2. Restart Claude Code session to reinitialize MCP connections

Stale pages from a previous session

If list_pages shows pages that aren't yours (leftover from a crashed session), close them before starting your work:

close_page(pageId: "<stale-page-id>")

Navigation timeout

Some pages take a long time to load. Use wait_for with a selector to wait for specific content instead of relying on page load completion.

Can't find an element

Use take_snapshot to see the accessibility tree structure, then target elements by their accessible name or role.

Tips

• Prefer take_snapshot over take_screenshot when you need to extract text content — it's faster and gives structured data
• Use take_screenshot when you need to see visual layout or show the user what the page looks like
• For SPAs, wait for content to render after navigation before taking snapshots
• If a page requires authentication, let the user know — headless browsers start with no session/cookies
• Each server instance has its own isolated browser profile, so cookies and state are not shared between instances
• The chrome-devtools-visible instance opens a real browser window — useful for demos or debugging but slower to start