chrome-devtools
npx skills add https://github.com/cygnusfear/agent-skills --skill chrome-devtools
Agent 安装分布
Skill 文档
Chrome DevTools MCP Skill
Control Chrome browser programmatically via chrome-devtools-mcp.
â¡ Performance Tip: Always use the patterns in “Quick Start” section below. They include cleanup and run everything in ONE command to avoid browser lock issues.
Speed Comparison:
- â OLD WAY: Multiple attempts, manual cleanup, 5-10 commands â ~60-90 seconds
- â NEW WAY: One optimized command â ~5-10 seconds
ð See also: MCP CLI Guide for general MCP CLI patterns and best practices
Setup
# macOS
brew tap f/mcptools
brew install mcp
# Windows/Linux
go install github.com/f/mcptools/cmd/mcptools@latest
Quick Start (FASTEST)
Check Console Errors on localhost
Single command (copy-paste ready):
pkill -9 -f "chrome-devtools-mcp" 2>/dev/null; sleep 1; echo -e 'navigate_page {"url":"http://localhost:3000"}\nlist_console_messages {"pageIdx":0}\nexit' | timeout 30 mcp shell bunx -y chrome-devtools-mcp@latest -- --isolated
Full Debug: Console + Network + Page Content
pkill -9 -f "chrome-devtools-mcp" 2>/dev/null; sleep 1; echo -e 'navigate_page {"url":"http://localhost:3000"}\nlist_console_messages {"pageIdx":0}\nlist_network_requests {"pageIdx":0}\ntake_snapshot {"verbose":true}\nexit' | timeout 30 mcp shell bunx -y chrome-devtools-mcp@latest -- --isolated
Screenshot a Page
pkill -9 -f "chrome-devtools-mcp" 2>/dev/null; sleep 1; echo -e 'navigate_page {"url":"https://example.com"}\ntake_screenshot {"fullPage":true,"format":"png"}\nexit' | timeout 30 mcp shell bunx -y chrome-devtools-mcp@latest -- --isolated
Execute JavaScript on Page
pkill -9 -f "chrome-devtools-mcp" 2>/dev/null; sleep 1; echo -e 'navigate_page {"url":"http://localhost:3000"}\nevaluate_script {"function":"() => document.querySelectorAll(\"div\").length"}\nexit' | timeout 30 mcp shell bunx -y chrome-devtools-mcp@latest -- --isolated
â¡ Pattern: cleanup; sleep; echo commands | timeout shell
This avoids:
- â Browser profile locks (cleanup first)
- â Multiple shell sessions (one pipeline)
- â Hanging commands (timeout wrapper)
Usage
List available tools:
mcp tools bunx -y chrome-devtools-mcp@latest
Call tools (individual commands):
# Navigate to a page (includes page list in response)
mcp call navigate_page bunx -y chrome-devtools-mcp@latest -- --isolated --params '{"url":"https://example.com"}'
# Take screenshot
mcp call take_screenshot bunx -y chrome-devtools-mcp@latest -- --isolated --params '{"fullPage":true,"format":"png"}'
# Take snapshot (page content as text)
mcp call take_snapshot bunx -y chrome-devtools-mcp@latest -- --isolated --params '{"verbose":false}'
# List console messages
mcp call list_console_messages bunx -y chrome-devtools-mcp@latest -- --isolated --params '{"pageIdx":0}'
# Execute JavaScript
mcp call evaluate_script bunx -y chrome-devtools-mcp@latest -- --isolated --params '{"function":"() => document.title"}'
Interactive shell mode (RECOMMENDED – maintains single browser instance):
mcp shell bunx -y chrome-devtools-mcp@latest -- --isolated
# Then run commands:
navigate_page {"url":"https://example.com"}
take_snapshot {"verbose":false}
list_console_messages {"pageIdx":0}
new_page {"url":"https://httpbin.org"}
exit
With Chrome launch options:
# Headless mode
mcp call navigate_page bunx -y chrome-devtools-mcp@latest -- --headless --isolated --params '{"url":"https://example.com"}'
# Custom viewport
mcp call navigate_page bunx -y chrome-devtools-mcp@latest -- --viewport "1920x1080" --isolated --params '{"url":"https://example.com"}'
# Connect to existing Chrome instance (no --isolated needed)
mcp call navigate_page bunx -y chrome-devtools-mcp@latest -- --browserUrl http://127.0.0.1:9222 --params '{"url":"https://example.com"}'
Common Workflows
Check localhost console logs:
# Shell mode (RECOMMENDED):
mcp shell bunx -y chrome-devtools-mcp@latest -- --isolated
# Then: navigate_page {"url":"http://localhost:3000"}
# Then: list_console_messages {"pageIdx":0}
# Or as individual commands:
mcp call navigate_page bunx -y chrome-devtools-mcp@latest -- --isolated --params '{"url":"http://localhost:3000"}'
mcp call list_console_messages bunx -y chrome-devtools-mcp@latest -- --isolated --params '{"pageIdx":0}'
Automate form filling:
mcp shell bunx -y chrome-devtools-mcp@latest -- --isolated
# Then in shell:
# navigate_page {"url":"https://example.com/login"}
# take_snapshot {"verbose":false} # Get UIDs for elements
# fill {"uid":"#username","value":"user"}
# fill {"uid":"#password","value":"pass"}
# click {"uid":"#submit"}
Network debugging:
# Shell mode (RECOMMENDED):
mcp shell bunx -y chrome-devtools-mcp@latest -- --isolated
# Then: navigate_page {"url":"https://example.com"}
# Then: list_network_requests {"pageIdx":0}
# Or as individual commands:
mcp call navigate_page bunx -y chrome-devtools-mcp@latest -- --isolated --params '{"url":"https://example.com"}'
mcp call list_network_requests bunx -y chrome-devtools-mcp@latest -- --isolated --params '{"pageIdx":0}'
Key Tools
- navigate_page – Go to URL (returns page list)
- new_page – Open new tab (returns page list)
- select_page – Switch to different page by index
- close_page – Close page by index
- click – Click element
- fill – Fill input/textarea
- evaluate_script – Run JavaScript
- take_screenshot – Capture page as image
- take_snapshot – Get page content as text with UIDs
- list_console_messages – View console logs (use
{"pageIdx":0}) - list_network_requests – View network activity (use
{"pageIdx":0}) - wait_for – Wait for text/condition
Important Notes
- Always use
bunx(notnpx) for chrome-devtools-mcp - Correct syntax:
mcp call TOOL_NAME bunx -y chrome-devtools-mcp@latest -- --isolated --params '{...}'- Server command comes AFTER tool name, BEFORE –params
- Parameters are passed directly (no “arguments” wrapper)
- Use
-- --isolatedto create temporary profile (prevents browser lock conflicts)
- Server exposes full browser content – avoid sensitive data
Working with Multiple Commands
Problem: Each mcp call starts a new server & browser instance, causing conflicts.
Solutions:
-
Shell mode (RECOMMENDED): Maintains single browser instance
mcp shell bunx -y chrome-devtools-mcp@latest -- --isolated -
Individual calls: Use
-- --isolatedfor one-off commandsmcp call navigate_page bunx -y chrome-devtools-mcp@latest -- --isolated --params '{"url":"..."}' -
Connect to running Chrome: Start Chrome with debugging enabled
/Applications/Google\ Chrome.app/Contents/MacOS/Google\ Chrome --remote-debugging-port=9222 mcp call navigate_page bunx -y chrome-devtools-mcp@latest -- --browserUrl http://127.0.0.1:9222 --params '{"url":"..."}'
Known Issues (mcp CLI v0.7.1)
â ï¸ Bug: Tools with empty parameter schemas fail with “Invalid arguments” error.
Workaround: Provide at least one optional parameter:
- â
list_console_messages {"pageIdx":0}– Works - â
list_network_requests {"pageIdx":0}– Works - â
take_snapshot {"verbose":false}– Works - â
list_pages– Broken, usenavigate_pageornew_pageinstead (they return page list)
Troubleshooting
Problem: “The browser is already running” error
This happens when a previous Chrome instance is still holding the profile lock.
Quick fix (ALWAYS DO THIS FIRST):
pkill -9 -f "chrome-devtools-mcp" 2>/dev/null; sleep 1
Then run your command:
echo -e 'navigate_page {"url":"YOUR_URL"}\nlist_console_messages {"pageIdx":0}\nexit' | mcp shell bunx -y chrome-devtools-mcp@latest -- --isolated
Problem: Multiple failed attempts
Don’t create multiple shell sessions – keep ONE session open and run multiple commands:
â BAD (slow – 2 separate sessions):
echo 'navigate_page {"url":"..."}' | mcp shell bunx -y chrome-devtools-mcp@latest -- --isolated
echo 'list_console_messages {"pageIdx":0}' | mcp shell bunx -y chrome-devtools-mcp@latest -- --isolated
â GOOD (fast – 1 session, multiple commands):
echo -e 'navigate_page {"url":"..."}\nlist_console_messages {"pageIdx":0}\nexit' | mcp shell bunx -y chrome-devtools-mcp@latest -- --isolated
Problem: Commands hanging/timing out
Use timeout to prevent hanging:
timeout 30 mcp shell bunx -y chrome-devtools-mcp@latest -- --isolated