playwright-mcp-orchestrator
A multi-agent orchestrator that enables multiple AI agents to share a single Chrome browser instance with tab isolation and shared state. It preserves local browser data like cookies and logins while allowing concurrent agent sessions through the official Playwright MCP.
README
playwright-mcp-orchestrator
Multi-agent orchestrator for Playwright MCP — multiple AI agents sharing one Chrome browser with tab isolation. Local-first, preserving your real browser state (cookies, logins, extensions, localStorage).
The Problem
The official @playwright/mcp only supports a single agent per browser. When multiple Claude Code agents (or any MCP clients) try to use it simultaneously, they fight over the same tabs and interfere with each other.
Existing workarounds all sacrifice something:
--isolatedmode gives each agent a fresh profile (no saved logins/cookies)- Separate
--user-data-dirper agent means separate profiles (no shared state) - Community alternatives (
ultimate-playwright-mcp,concurrent-browser-mcp) are unmaintained or don't share profiles
The Approach
Use a single Chrome instance with CDP (Chrome DevTools Protocol) as the shared backend, and orchestrate multiple @playwright/mcp sessions on top — each scoped to its own tabs.
┌─────────────────────────────────────────────┐
│ Chrome (one instance) │
│ --remote-debugging-port=9222 │
│ --user-data-dir=<profile> │
│ │
│ ┌─────────┐ ┌─────────┐ ┌─────────┐ │
│ │ Agent 1 │ │ Agent 2 │ │ Agent 3 │ tabs │
│ │ tabs │ │ tabs │ │ tabs │ │
│ └────┬────┘ └────┬────┘ └────┬────┘ │
└───────┼──────────┼──────────┼───────────────┘
│ │ │ CDP (ws://127.0.0.1:9222)
┌───────┼──────────┼──────────┼───────────────┐
│ ▼ ▼ ▼ │
│ ┌────────────────────────────────┐ │
│ │ playwright-mcp-orchestrator │ │
│ │ │ │
│ │ - Tab group isolation │ │
│ │ - Session routing │ │
│ │ - Uses official @playwright/mcp│ │
│ └────────────────────────────────┘ │
└─────────────────────────────────────────────┘
Key design decisions
- Delegates to official
@playwright/mcp— not a reimplementation. New tools and features from Microsoft are available automatically. - Shared user profile — all agents see the same cookies, logins, localStorage.
- Tab isolation — each agent operates on its own tabs, no interference.
- CDP connection — connects to a running Chrome instance via
--remote-debugging-port.
Status
Early development. The CDP foundation has been validated:
- Chrome 145 with
--remote-debugging-port=9222+--user-data-dirworks - Multiple
playwright-coreconnections viaconnectOverCDPto the same browser work - Each connection can create and manage its own tabs independently
- All connections share the same BrowserContext (cookies, logins, storage)
Chrome Setup
Chrome 136+ silently ignores --remote-debugging-port on the default profile directory. You must use --user-data-dir pointing to a non-default path.
Copy your profile (preserves all logins)
# Quit Chrome first
osascript -e 'tell application "Google Chrome" to quit'
# Copy profile
cp -r ~/Library/Application\ Support/Google/Chrome /tmp/chrome-cdp-profile
# Launch with CDP
open -a "Google Chrome" --args \
--user-data-dir=/tmp/chrome-cdp-profile \
--remote-debugging-port=9222
Verify CDP is working
curl -s http://127.0.0.1:9222/json/version
Research
Why not existing solutions?
| Project | Stars | Issue |
|---|---|---|
@playwright/mcp (official) |
— | Single agent only. Issues #893, #1294 open. |
ultimate-playwright-mcp |
1 | Full reimplementation, abandoned after 1 day. Can't track official updates. |
concurrent-browser-mcp |
— | Isolated ephemeral instances, no shared profile. |
playwright-parallel-mcp |
— | Separate processes, separate profiles. |
mcp-playwright-cdp |
47 | CDP but no tab isolation. Abandoned. |
Why not other browser frameworks?
| Framework | Stars | Issue for our use case |
|---|---|---|
| Browser Use | 80k | Single-session locally. Cloud for concurrency. |
| Stagehand/Browserbase | 21k | Cloud-only shared state. Not local. |
| Agent Browser (Vercel) | 21k | CLI, not MCP. No tab isolation. CDP still maturing. |
| Dev Browser | 3.8k | Skill, not MCP. Same multi-agent gap (#33). |
| Skyvern | 21k | Vision-first, different paradigm. |
The multi-agent shared-profile problem is unsolved
No well-maintained project provides: real user profile + multiple concurrent agents + tab isolation + MCP interface. This project aims to be the first.
License
MIT
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。