cdp-chrome

name	cdp-chrome
description	Optional per-OS-user headed Chrome provider for browser automation. Use when an environment chooses cdp-chrome for visible GUI Chrome: social media, JS-rendered pages, logged-in sites, anti-bot pages, forms, screenshots, and live site inspection. Not required when an equivalent provider exists.

CDP Chrome: Per-User Headed Browser Provider

Scope: Optional implementation of the abstract headed-browser capability. Do not force this plugin when the user already has an equivalent provider.

Why This Exists

chrome-devtools-mcp can launch Chrome with automation flags such as --enable-automation, which sets navigator.webdriver = true. This plugin instead connects MCP to a normal GUI Chrome process with a persistent profile.

The process is shared across agents for the same OS user only. On multi-user machines, every OS user should configure a different port/profile so agents fail fast instead of connecting to another user's Chrome.

Config

steroids config file:

• macOS/Linux: ~/.config/steroids.json
• Windows: %APPDATA%\steroids.json
• shell form: ${APPDATA:-$HOME/.config}/steroids.json

Default config:

{ "cdp-chrome": { "port": 9224, "profile_dir": "~/.config/cdp-chrome/profile" } }

Existing configs with only port still work; profile_dir defaults to ~/.config/cdp-chrome/profile. Do not configure a shared download directory; Chrome default downloads are left alone.

Setup / Doctor

1. Choose a unique cdp-chrome.port and profile_dir for this OS user in the steroids config file.

2. Run the plugin script:

   plugins/chrome/skills/cdp-chrome/scripts/doctor.sh
   

   It verifies current-user config, port ownership, and Chrome profile consistency. If it reports another OS user or another profile on the port, choose a different cdp-chrome.port and rerun it.

3. Start Chrome:

   plugins/chrome/skills/cdp-chrome/scripts/start.sh
   

   The script creates profile_dir, refuses occupied/wrong-user/wrong-profile listeners on macOS, and starts normal GUI Chrome without --enable-automation.

4. Log in manually to needed sites. Sessions persist in the configured profile.

MCP Registration

Claude Code and Codex: installing the chrome plugin provides cdp-chrome through plugin-local .mcp.json. Claude Code documents ${CLAUDE_PLUGIN_ROOT} for plugin MCP paths; current Codex plugin loading has been verified to start plugin MCP entries with cwd: "." at the installed plugin root. The shared .mcp.json uses a small shell launcher to support both cases, then runs skills/cdp-chrome/scripts/mcp-launcher.sh; that launcher reads the current user's config, validates an existing listener when possible, then execs:

npx -y chrome-devtools-mcp@latest --browserUrl http://127.0.0.1:<port> --no-usage-statistics \
  --no-category-performance --no-category-emulation --no-category-network

Hermes: plugin-local MCP config is not auto-loaded. Register an equivalent mcp_servers.cdp-chrome manually and point it at this plugin's mcp-launcher.sh or at the same chrome-devtools-mcp command with your configured port. Restart/reload MCP after config changes.

Do not create a duplicate project-level ./.mcp.json for the same server; duplicate MCP definitions can connect to different ports.

Target Binding

This skill is only satisfied when the agent is operating on the configured shared CDP endpoint. Similar-looking tools such as mcp__chrome_devtools__*, Playwright, Puppeteer, or browser-use are not substitutes unless they are explicitly registered as the cdp-chrome server for this plugin and proven to use the configured http://127.0.0.1:<port> endpoint. They can silently attach to a different browser or target, even when their API is based on Chrome DevTools.

If the expected cdp-chrome MCP namespace is not exposed, do not guess with another browser tool. First run doctor.sh, then use the configured endpoint directly:

curl -s "http://127.0.0.1:<port>/json/list"

Pick the intended target from /json/list and operate through its webSocketDebuggerUrl, or report that the cdp-chrome MCP namespace is missing. A page list from any other tool is not proof that this skill is attached to the shared Chrome.

Profile-scoped Validation

CDP Chrome 使用独立的 profile_dir。在这个实例里看到的 cookie、扩展、Preferences、WebRTC 或 DNS 结果只证明该 profile；doctor.sh 也只证明端口、进程与 user-data-dir 绑定正确，不能证明用户日常 Chrome 已修复。

涉及浏览器 policy、WebRTC、DNS、扩展或登录态的结论时，先记录当前 binary、user-data-dir 与 chrome://version 的 Profile Path，并明确结论的作用域。profile preference 不得外推到其它 profile；只有 chrome://policy 中 Source=Platform、Level=Mandatory、Status=OK 的受管策略可以跨 profile，但仍要在实际出问题的浏览器和新开的无痕窗口复测。保持只读验收，不用 CDP Chrome 顺手修改 profile 设置。

Agent Rules

1. Use only MCP tools from server name cdp-chrome (mcp__cdp-chrome__* in Claude/Codex, mcp_cdp_chrome_* style in Hermes), or the direct configured CDP endpoint fallback above. Do not fall back to other Chrome/Playwright/Puppeteer MCP tools; they may launch automated Chrome or attach to a different Chrome target.
2. Never launch your own Chrome. Use start.sh if the configured instance is not running.
3. Before browser work, run doctor.sh when setup changed or when connection errors occur.
4. Open tabs for your task and close them when done. Do not touch other agents' tabs.
5. Do not clear cookies, change profile settings, install extensions, or modify the browser profile.
6. Parallel agents should run in separate agent processes. A single MCP process can have global selected-page state even though Chrome tabs have independent CDP target IDs.
7. Understand pages visually first. Before interacting with a complex or unknown page, call take_screenshot (~800–1,600 vision tokens) to see the layout. Do NOT call take_snapshot for this purpose — its A11Y text tree costs 10K–540K chars (2.5K–135K text tokens) on complex pages and often exceeds tool limits. After the screenshot gives you spatial understanding, use evaluate_script for precise extraction/action.
8. Reserve take_snapshot for simple pages only. Login forms, settings panels, confirmation dialogs — pages where the A11Y tree is expected to be < 5K chars. For anything else, screenshot + evaluate_script is both cheaper and more effective.
9. Cap evaluate_script results. When writing extraction JS, truncate or paginate output in-script (e.g. .slice(0, 100) for arrays, .slice(0, 8000) for text). Do not return unbounded DOM content or full page text.

Quick Checks

PORT=$(python3 - <<'PY'
import json, os
p=os.path.join(os.environ.get('APPDATA', os.path.join(os.environ['HOME'], '.config')), 'steroids.json')
try:
    print(json.load(open(os.path.expanduser(os.path.expandvars(p)))).get('cdp-chrome', {}).get('port', 9224))
except FileNotFoundError:
    print(9224)
PY
)
curl -s "http://127.0.0.1:$PORT/json/version"
curl -s "http://127.0.0.1:$PORT/json/list"

Red flags: another OS user owns the port, process args lack the configured --user-data-dir, --enable-automation, --remote-debugging-pipe, temp puppeteer_dev_chrome_profile-*, or unexpected logouts. Stop and fix config/MCP registration.

Page Interaction

See references/page-interaction.md for detailed patterns and examples.

Tool selection guide:

目的	工具	Token 成本	说明
看懂页面布局	take_screenshot	~800–1,600 vision tokens	复杂/未知页面首选
精准提取/操作	evaluate_script	~650 text tokens（可控）	主力工具
获取元素 UID	take_snapshot	2.5K–135K text tokens	仅限简单页面
导航	navigate_page	~190 text tokens	—