ChatGPT CDP MCP
MCP server that turns one authenticated ChatGPT browser page into a bounded MCP server, exposing tools for creating, listing, searching, exporting, sending, editing chats and downloading media via a CDP driver.
README
ChatGPT CDP MCP
Русский · Quick start · Driver guide · Tool reference

Turn one authenticated ChatGPT browser page into a bounded MCP server — with a driver you control and a single-page lease you can verify.
ChatGPT CDP MCP gives an MCP client a clean, typed tool surface for a browser page you already own. Bring any CDP-capable adapter (Chrome DevTools Protocol, Playwright, BrowserClaw, or your own), point the server at it, and keep the browser integration separate from the MCP contract.
- Seven MCP operations: create, list, search, export, send, edit, and download.
- One owned browser page per MCP process; every operation rechecks its lease.
- Existing chats are discoverable; consequential operations are restricted to one disposable fixture created by the server.
- Opaque refs keep browser-native IDs out of MCP responses.
- A bundled mock driver lets you verify the complete MCP wiring before connecting a real account.
Install
npm install --global git+https://github.com/megamen32/chatgpt-cdp-mcp.git
Requires Node.js 20+ and a local driver module. The command installs the server; a driver is the unavoidable connection to your already authenticated browser.
Start in minutes
Try the safe local demo first — it does not open a browser or contact ChatGPT:
git clone https://github.com/megamen32/chatgpt-cdp-mcp.git && cd chatgpt-cdp-mcp && npm ci && CDP_CHAT_DRIVER_MODULE=./examples/mock-driver.mjs npm start
Then configure your MCP client:
{
"mcpServers": {
"chatgpt": {
"command": "chatgpt-cdp-mcp",
"env": {
"CDP_CHAT_DRIVER_MODULE": "/absolute/path/to/your-chatgpt-driver.mjs",
"CDP_CHAT_MEDIA_ROOT": "/absolute/path/to/private-media"
}
}
}
}
Build the real driver from the compact driver contract. The server intentionally does not ship a browser login flow, profile, cookie, or hosted endpoint.
What it exposes
| Tool | Purpose |
|---|---|
new_chat |
Creates one disposable fixture chat without submitting a prompt. |
list_chats |
Lists visible chats by unread, working, or recent. |
search_chat |
Searches visible titles and message text from one fresh snapshot. |
export_chat |
Exports a fixture chat as bounded JSON or Markdown. |
send_message |
Sends once to the fixture after explicit confirmation. |
edit_message |
Edits a fixture message with an optimistic guard. |
download_media |
Saves an allowlisted fixture attachment under a confined directory. |
Read the full tool reference for inputs, limits, and expected results.
Why the fixture boundary?
The most expensive browser failures happen when an automation targets the wrong conversation. This project makes read-only discovery convenient, then requires new_chat to create the only mutable/exportable target. That makes integration testing fast without turning a production conversation into a test fixture.
Project layout
MCP client
│ stdio
▼
ChatGPT CDP MCP
│ CdpChatDriver
▼
one browser page you own
The public package owns the tool contract, opaque references, bounds, and fixture lifecycle. Your adapter owns browser automation and authentication. See architecture.
Development
npm ci
npm test
The test suite uses an in-memory browser page; no ChatGPT account, browser profile, or network access is required.
Important notes
- Use only a browser page and account you are authorized to control, and comply with the applicable terms of service.
- This is an independent open-source project; it is not affiliated with or endorsed by OpenAI.
- The server never includes a browser profile, cookie, token, or account data in the repository.
Learn more
- Quick start
- Writing a real CDP driver
- Tool reference
- Architecture
- Configuration
- Security model
- Contributing
MIT © 2026 megamen32
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。
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。
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。
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。