chrome-local-mcp
Lightweight MCP server for controlling a Windows Chrome instance from WSL, designed for AI coding agents with minimal token overhead and human-in-the-loop support.
README
chrome-local-mcp
English | 简体中文
Lightweight MCP server for controlling a Windows Chrome instance from WSL. Designed for AI coding agents (Claude Code, Codex, Cursor) that need browser interaction with minimal token overhead and human-in-the-loop support.
Why?
Existing browser MCP tools (Chrome DevTools MCP, Playwright MCP) have three problems in WSL environments:
| Problem | This project's solution |
|---|---|
| Schema overhead: 13,700-18,000 tokens before any interaction | ~1,000 tokens for 17 tools |
| Long page content floods LLM context (accessibility tree up to 124K tokens) | Long content auto-saved to file, only summary returned |
| No human intervention mechanism for CAPTCHAs/verification | Built-in auto-detection + wait_for_human tool |
Prerequisites
Windows Chrome running with remote debugging enabled:
chrome.exe --remote-debugging-port=9222
Important: Close all existing Chrome windows before running this command, otherwise the debug port won't open.
Or start an isolated Chrome profile without closing your normal browser:
.\scripts\start-chrome-debug.ps1
The startup script supports multiple modes:
# Recommended for screenshots: keeps rendering active when Chrome is covered/backgrounded.
.\scripts\start-chrome-debug.ps1 -Mode stable
# Plain visible Chrome with an isolated profile.
.\scripts\start-chrome-debug.ps1 -Mode interactive
# No visible Chrome window. Best for automation, but weaker for manual CAPTCHA/login flows.
.\scripts\start-chrome-debug.ps1 -Mode headless
Equivalent npm shortcuts are available on Windows:
npm run chrome:stable
npm run chrome:stable:wsl
npm run chrome:interactive
npm run chrome:headless
stable adds Chrome flags that reduce background throttling and native occlusion issues:
--disable-backgrounding-occluded-windows
--disable-renderer-backgrounding
--disable-background-timer-throttling
--disable-features=CalculateNativeWinOcclusion
Useful options:
.\scripts\start-chrome-debug.ps1 -Mode stable -Port 9333 -UserDataDir D:\tmp\chrome-mcp-profile
.\scripts\start-chrome-debug.ps1 -Mode stable -WindowWidth 1600 -WindowHeight 1000
.\scripts\start-chrome-debug.ps1 -Mode stable -ExtraArgs "--lang=en-US"
.\scripts\start-chrome-debug.ps1 -Mode stable -ReadyTimeoutSeconds 20
.\scripts\start-chrome-debug.ps1 -Mode stable -ReuseExisting
.\scripts\start-chrome-debug.ps1 -Mode stable -DryRun
By default, the script refuses to launch a second Chrome when the requested debug port is already listening. Use -ReuseExisting only when you intentionally want to attach to the currently running Chrome CDP endpoint.
If the MCP server runs in WSL and cannot reach Windows localhost:9222, bind Chrome to a reachable address and point the MCP server at the Windows host IP:
.\scripts\start-chrome-debug.ps1 -Mode stable -RemoteDebuggingAddress 0.0.0.0
Or use the npm shortcut:
npm run chrome:stable:wsl
Then start the MCP server with the Windows host IP instead of localhost:
node dist/index.js --host <windows-host-ip> --port 9222
0.0.0.0 exposes Chrome DevTools beyond local loopback. Use it only on a trusted local machine/network and prefer firewall rules that restrict access.
Verify Chrome DevTools Protocol is reachable:
curl http://localhost:9222/json/version
Install
cd /path/to/chrome-local-mcp
npm install
npm run build
Setup
Windows Codex CLI
codex mcp add chrome-local -- node "D:\repository\chrome-local-mcp\dist\index.js"
codex mcp list
If Chrome uses a non-default host or port:
codex mcp add chrome-local -- node "D:\repository\chrome-local-mcp\dist\index.js" --host localhost --port 9222
Windows Claude Code
claude mcp add --scope user chrome-local -- node "D:\repository\chrome-local-mcp\dist\index.js"
claude mcp list
Claude Code
claude mcp add --scope user chrome-local -- node /path/to/chrome-local-mcp/dist/index.js
OpenAI Codex CLI
Edit ~/.codex/config.toml:
[mcp_servers.chrome-local]
command = ["node", "/path/to/chrome-local-mcp/dist/index.js"]
Cursor / VS Code
Add to .cursor/mcp.json or VS Code MCP settings:
{
"mcpServers": {
"chrome-local": {
"command": "node",
"args": ["/path/to/chrome-local-mcp/dist/index.js"]
}
}
}
Custom host/port
If WSL and Windows don't share localhost (WSL2 NAT mode), pass the Windows host IP:
node dist/index.js --host 172.x.x.1 --port 9222
You can also configure defaults with environment variables:
| Variable | Description | Default |
|---|---|---|
CHROME_HOST |
Chrome DevTools Protocol host | localhost |
CHROME_PORT |
Chrome DevTools Protocol port | 9222 |
CHROME_MCP_TMPDIR |
Directory for saved content and screenshots | OS temp directory + chrome-mcp |
CHROME_MCP_CHALLENGE_LOG |
JSONL log path for human verification events detected during normal MCP use | repo logs/challenge-events.jsonl |
CHROME_MCP_CHALLENGE_LOG_DEDUPE_MS |
Dedupe window for the same URL/challenge log event | 60000 |
CHROME_MCP_DISABLE_CHALLENGE_LOG |
Set to 1 to disable challenge event logging |
unset |
Tools
Navigation
| Tool | Description |
|---|---|
navigate(url) |
Open a URL. Auto-detects verification challenges. |
wait_for(selector, timeout?) |
Wait for an element to appear. |
go_back() |
Browser back button. |
Content
| Tool | Description |
|---|---|
get_content(selector?, max_length?) |
Extract page text. Short text returned directly; long text saved to file with preview. |
evaluate(expression) |
Execute JavaScript in page context. Long results are saved to file with preview. |
screenshot(path?, full_page?) |
Save screenshot to file (never inlined into context). |
Interaction
| Tool | Description |
|---|---|
click(selector) |
Click an element. Auto-detects verification after click. |
fill(selector, value) |
Type into a form field. |
press_key(key) |
Press a key (Enter, Tab, Escape, etc.). |
scroll(direction, amount?) |
Scroll up or down. |
Tabs
| Tool | Description |
|---|---|
tab_list() |
List all open tabs. |
tab_open(url) |
Open URL in a new tab. |
tab_switch(id) |
Switch to a tab by ID. |
tab_close(id?) |
Close a tab. |
Search
| Tool | Description |
|---|---|
search(query, max_results?) |
Google search with structured results. Auto-detects CAPTCHAs. |
Human Intervention
| Tool | Description |
|---|---|
wait_for_human(reason, wait_until_gone?, wait_until_present?, timeout?) |
Pause for human action (CAPTCHA, login, etc.). Polls until condition is met. |
check_page_status() |
Detect verification challenges on current page. |
Human-in-the-loop
Verification challenges are auto-detected after every navigation-related action. When detected, the tool response includes:
{
"status": "human_verification_needed",
"challenges": ["cloudflare", "turnstile"],
"message": "Page has a verification challenge. Complete it in the browser, then call wait_for_human or retry."
}
Detected challenge types: cloudflare, recaptcha, hcaptcha, turnstile, google_sorry, google_consent, age_gate.
Access prompts that do not block readable content, such as article subscription prompts or paid preview banners, should not return human_verification_needed.
When normal MCP usage detects a real human verification challenge, it appends a JSONL event to logs/challenge-events.jsonl by default:
{
"event": "human_verification_detected",
"triggered": true,
"triggered_at": "2026-05-02T02:10:00.000Z",
"webpage": {
"url": "https://example.com/",
"title": "Example"
},
"has_challenge": true,
"challenges": ["cloudflare"]
}
Use CHROME_MCP_CHALLENGE_LOG to write this log somewhere else.
Challenge regression log
Run the canary monitor against a Chrome instance to record pages where challenge detection may be too aggressive:
npm run build
npm run monitor:challenges
The monitor writes JSONL entries to logs/challenge-regressions.jsonl. Each entry includes checked_at, triggered_at, and the webpage URL/title. By default it only logs; use -- --fail-on-issue true if you want issues to make the command exit non-zero.
Typical flow
Agent: navigate("https://protected-site.com")
-> {"status": "human_verification_needed", "challenges": ["cloudflare"]}
Agent: wait_for_human(reason="Cloudflare verification", wait_until_gone="#challenge-running")
-> User completes verification in Chrome
-> {"status": "ready", "waited_ms": 8000}
Agent: get_content()
-> Page content extracted normally
Token efficiency
| Metric | chrome-local-mcp | Chrome DevTools MCP | Playwright MCP |
|---|---|---|---|
| Schema overhead | ~1,000 tokens | ~18,000 tokens | ~13,700 tokens |
| navigate response | ~60 tokens | varies | ~3,800+ tokens (accessibility tree) |
| screenshot | ~30 tokens (file path) | ~4,000 tokens (base64) | ~4,000 tokens |
| 10-page research task | ~5,000 tokens | ~70,000+ tokens | ~500,000+ tokens |
These numbers are order-of-magnitude estimates. Actual token usage depends on the client, tool schema version, page complexity, and task flow.
Architecture
Claude Code / Codex (WSL)
↓ stdio
chrome-local-mcp (Node.js, WSL)
↓ CDP WebSocket (ws://localhost:9222)
Chrome (Windows, with --remote-debugging-port=9222)
- Transport: stdio (MCP standard)
- CDP client:
chrome-remote-interface - Long content saved to
/tmp/chrome-mcp/; agents read files on demand
Development
npm run build # Compile TypeScript
npm run dev # Watch mode
npm run test # Run test suite
License
MIT
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.