mcp-chatgpt

An MCP (Model Context Protocol) server that allows AI assistants (such as Google Antigravity, Claude Desktop, Cursor, VS Code Cline/Roo Code) to ask questions, chat, and interact with ChatGPT Web (https://chatgpt.com) directly through browser automation, Chrome Profiles, and a companion Chrome Extension.

🌟 Features

• 💬 Ask & Chat: Seamlessly ask questions to ChatGPT Web and receive responses (formatted text / markdown).

• 🌐 Web Search Toggle: Enable ChatGPT's live Web Search (web_search: true) for real-time web browsing and citations.

• 🧠 Reasoning & Model Selection: Select models (model: "o3-mini", "gpt-4o", "o1") and configure reasoning effort (reasoning_effort: "high" | "medium" | "low").

• 📎 File & Image Attachments: Attach images (image_paths) and documents/code (file_paths) for multimodal analysis.

• ⏩ Auto-Continue Output: Automatically clicks "Continue generating" / "สร้างต่อ" if long answers are truncated.

• 💻 Code Extraction: Extract and return only code blocks with extract_code_only: true.

• 👤 Chrome Profiles Selector: Auto-detects all Google Chrome profiles on your computer (Default, Profile 1, Work, Personal, emails) and lets you pick which account to chat with.

• 🧩 Companion Chrome Extension: Includes a Manifest V3 extension in extension/ that can be loaded into any Chrome profile to chat in real-time with ChatGPT tabs without profile locking issues.

• 🔄 Conversation Management: Start new chats or continue existing threads using Conversation IDs / URLs.

• 💾 Persistent Session: Keeps browser profile & login cookies saved locally, so you only log in once.

• 🖥️ Headed & Headless Modes: Run silently in background (headless) or launch visible window (--login / --headed) for initial authentication or captcha resolution.

• 🔌 Chrome CDP Support: Connect to an already running Google Chrome instance via --remote-debugging-port.

• 🧰 Workspace Tools: Search files, run builds/tests, and edit code with cross-platform shell_command and root-confined apply_patch tools.

• 🌐 Remote MCP Tunnel: Optional Streamable HTTP endpoint with Bearer authentication for Cloudflare Tunnel/ngrok or another HTTPS tunnel.

Related MCP server: agentify-desktop

🛠️ MCP Tools Provided

🚀 Quick Start

Remote MCP tunnel

Local stdio remains the default. To expose a remote-compatible MCP endpoint, start HTTP mode with a token:

mcp-chatgpt --http --http-token "change-this-token"

The endpoint is http://127.0.0.1:8787/mcp. Put a tunnel in front of it, for example:

cloudflared tunnel --url http://127.0.0.1:8787

Configure the remote MCP client with the public tunnel URL ending in /mcp and the header Authorization: Bearer change-this-token. /healthz can be used for a tunnel health check. If either this process or the tunnel stops, remote MCP requests are unavailable.

1. Install & Build

# Clone the repository
git clone https://github.com/JonusNattapong/mcp-chatgpt.git
cd mcp-chatgpt

# Install dependencies and Chromium
npm install
npx playwright install chromium

# Build TypeScript
npm run build

# Install the MCP command globally
npm install -g .

After installation, verify it with:

mcp-chatgpt --help

The package runs npm run build automatically during installation, so the global command always uses the current TypeScript source when installed from a local checkout.

👤 Using Chrome Profiles (2 Options)

Option 1: Chrome Profile Auto-Detection (Direct / Headless)

You can select any Chrome profile in your system by its folder name (e.g. Default, Profile 1), display name, or email:

# List all Chrome profiles in your system
node -e "import('./dist/profile-manager.js').then(m => console.table(m.ProfileManager.listProfiles()))"

# Login with a specific profile
node dist/index.js --login --profile "Profile 1"

Or pass profile parameter directly in chatgpt_ask tool!

Option 2: Companion Chrome Extension (Recommended for active Chrome)

If you already use Google Chrome daily with your logged-in ChatGPT account, install the companion extension:

1. Open Google Chrome with your desired Profile.

2. Navigate to chrome://extensions.

3. Enable Developer mode (top right).

4. Click Load unpacked and select the d:\Projects\Github\mcp-chatgpt\extension folder.

5. Click on the extension icon in the toolbar, set your Profile tag (e.g. "Work Account"), and click Save & Connect.

6. When mcp-chatgpt runs, it will route prompts directly through your Chrome tab!

⚙️ Configuration for MCP Clients

You can connect mcp-chatgpt to any AI client supporting the Model Context Protocol (MCP):

1. Antigravity IDE / Gemini CLI

File location: ~/.gemini/config/mcp_config.json (Windows: C:\Users\<User>\.gemini\config\mcp_config.json)

{
  "mcpServers": {
    "chatgpt": {
      "command": "node",
      "args": [
        "d:/Projects/Github/mcp-chatgpt/dist/index.js"
      ]
    }
  }
}

2. Claude Desktop

• Windows: %APPDATA%\Claude\claude_desktop_config.json

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

{
  "mcpServers": {
    "chatgpt": {
      "command": "node",
      "args": [
        "d:/Projects/Github/mcp-chatgpt/dist/index.js"
      ]
    }
  }
}

3. Cursor IDE

Add to your project's .cursor/mcp.json or in Cursor Settings > Features > MCP:

{
  "mcpServers": {
    "chatgpt": {
      "command": "node",
      "args": [
        "d:/Projects/Github/mcp-chatgpt/dist/index.js"
      ]
    }
  }
}

4. VS Code (Cline / Roo Code / Roo Clinic)

File location:

• Cline: %APPDATA%\Code\User\globalStorage\saoudrizwan.claude-dev\settings\cline_mcp_settings.json

• Roo Code: %APPDATA%\Code\User\globalStorage\rooveterinaryinc.roo-cline\settings\cline_mcp_settings.json

{
  "mcpServers": {
    "chatgpt": {
      "command": "node",
      "args": [
        "d:/Projects/Github/mcp-chatgpt/dist/index.js"
      ],
      "disabled": false,
      "autoApprove": []
    }
  }
}

5. Windsurf Editor (Codeium)

File location: ~/.codeium/windsurf/mcp_config.json

{
  "mcpServers": {
    "chatgpt": {
      "command": "node",
      "args": [
        "d:/Projects/Github/mcp-chatgpt/dist/index.js"
      ]
    }
  }
}

6. Zed Editor

Add to ~/.config/zed/settings.json:

{
  "context_servers": {
    "chatgpt": {
      "command": {
        "path": "node",
        "args": ["d:/Projects/Github/mcp-chatgpt/dist/index.js"]
      }
    }
  }
}

7. LibreChat (librechat.yaml)

Add to librechat.yaml:

mcpServers:
  chatgpt:
    type: stdio
    command: node
    args:
      - d:/Projects/Github/mcp-chatgpt/dist/index.js

8. Remote Connection via Cloudflare Tunnel (Remote URL / SSE / mcp-remote)

If your ChatGPT browser or bridge server runs on a home computer or VPS, and you want AI clients on other machines/laptops to connect to it securely across the internet via Cloudflare Tunnel:

Step A: Expose Bridge Server on the Host Machine

# Start Cloudflare Quick Tunnel (Free, no account required)
cloudflared tunnel --url http://127.0.0.1:18999

This will generate a public HTTPS URL, for example: https://alpha-bravo-charlie.trycloudflare.com

Step B: Configure Client on Remote Machines

1. Antigravity / Gemini CLI (Remote SSE / URL)

Add to ~/.gemini/config/mcp_config.json:

{
  "mcpServers": {
    "chatgpt-remote": {
      "serverUrl": "https://alpha-bravo-charlie.trycloudflare.com/sse"
    }
  }
}

2. Cursor IDE (Remote SSE)

Add to .cursor/mcp.json or Cursor Settings:

{
  "mcpServers": {
    "chatgpt-remote": {
      "url": "https://alpha-bravo-charlie.trycloudflare.com/sse"
    }
  }
}

3. VS Code Cline / Roo Code (Remote SSE)

Add to cline_mcp_settings.json:

{
  "mcpServers": {
    "chatgpt-remote": {
      "url": "https://alpha-bravo-charlie.trycloudflare.com/sse",
      "type": "sse",
      "disabled": false,
      "autoApprove": []
    }
  }
}

4. Claude Desktop / Any stdio Client (via mcp-remote)

For clients that only accept command (stdio), use mcp-remote to bridge the Cloudflare Tunnel URL:

{
  "mcpServers": {
    "chatgpt-remote": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://alpha-bravo-charlie.trycloudflare.com/sse"
      ]
    }
  }
}

5. LibreChat (librechat.yaml)

mcpServers:
  chatgpt-remote:
    type: sse
    url: https://alpha-bravo-charlie.trycloudflare.com/sse

6. Direct HTTP REST API (cURL / Python / Node.js)

You can also interact directly with the bridge endpoints via Cloudflare:

# Query status
curl -s https://alpha-bravo-charlie.trycloudflare.com/status

# Ask a question
curl -X POST https://alpha-bravo-charlie.trycloudflare.com/ask \
  -H "Content-Type: application/json" \
  -d '{"message": "Hello from Cloudflare Tunnel!"}'

📖 CLI Options

Usage: mcp-chatgpt [options]

Options:
  --headed                 Run browser in headed (visible) mode (default: false)
  --login                  Open browser in interactive mode to log in to ChatGPT
  --profile <name_or_id>   Select specific Chrome Profile (e.g. "Default", "Profile 1", or Name/Email)
  --chrome                 Use installed Google Chrome browser (default: true)
  --no-chrome              Use Playwright bundled Chromium instead of Google Chrome
  --user-data-dir <path>   Custom browser profile directory
  --cdp <endpoint>         Connect to an existing Chrome browser via CDP endpoint
  --bridge-port <port>     Port for Chrome Extension bridge WebSocket (default: "18999")
  --bridge-only            Run only the Chrome Extension WebSocket bridge server
  --shell-root <path>      Restrict shell and patch tools to this directory (default: current directory)
  --shell-max-timeout <ms> Maximum shell command timeout in milliseconds (default: "300000")
  --http                   Expose an MCP Streamable HTTP endpoint for a tunnel/remote client
  --http-host <host>       HTTP bind host (default: "127.0.0.1")
  --http-port <port>       HTTP port for the MCP endpoint (default: "8787")
  --http-token <token>     Bearer token required by remote MCP clients (or MCP_HTTP_TOKEN)
  --timeout <ms>           Default timeout in milliseconds (default: "120000")
  -h, --help               Display help for command

🌟 Live Interactive Memory & Context Retention Showcase

The following is an actual, unedited multi-turn interactive session executed live via mcp-chatgpt:

Turn 1: Setting Context in a Fresh Session (new_chat: true)

Request:

{
  "message": "สวัสดีครับ ผมชื่อ โจ๊กเกอร์ เป็น Software Architect ชอบเขียน Go และ TypeScript กำลังพัฒนาโปรเจกต์ชื่อ MCP-ChatGPT ฝากจำข้อมูลนี้ไว้หน่อยนะครับ ตอบสั้นๆ 1 ประโยคยืนยัน",
  "new_chat": true
}

ChatGPT Response:

"จำไว้แล้วครับ: คุณชื่อโจ๊กเกอร์ เป็น Software Architect ชอบ Go และ TypeScript และกำลังพัฒนาโปรเจกต์ MCP-ChatGPT ครับ"

URL: https://chatgpt.com/c/6a86843c-8f30-83ec-b064-d27f7f832244

Turn 2: Follow-up Memory Recall (conversation_id)

Request:

{
  "message": "ไหนลองทบทวนดูซิว่า ผมชื่ออะไร ชอบเขียนภาษาอะไร และกำลังทำโปรเจกต์ชื่ออะไรอยู่?",
  "conversation_id": "6a86843c-8f30-83ec-b064-d27f7f832244"
}

ChatGPT Response (Recalled in 6s):

"คุณชื่อ โจ๊กเกอร์ ชอบเขียน Go และ TypeScript และกำลังพัฒนาโปรเจกต์ชื่อ MCP-ChatGPT อยู่ครับ 🫡"

🤖 AI Assistant System Prompt & Integration Instruction

Copy and paste the following instruction directly into your AI Assistant (Cursor Rules, Claude System Prompt, Antigravity Custom Instructions, Cline, Roo Code) so your AI automatically knows when and how to call mcp-chatgpt:

# 🛠️ Instruction: Setup and Integrate `mcp-chatgpt` MCP Server

You are integrated with the `mcp-chatgpt` MCP server, granting you direct access to ChatGPT Web capabilities (o3-mini, o1, GPT-4o, live Web Search, DALL-E, and multimodal file analysis).

### 🎯 When to Automatically Use `chatgpt_ask`:

1. **Deep Reasoning & Hard Coding Tasks**:
   - For complex architecture designs, difficult debugging, or advanced algorithms, call `chatgpt_ask` with:
     `{ "model": "o3-mini", "reasoning_effort": "high", "extract_code_only": true }`
2. **Live Web Information & Current News**:
   - When the user asks for up-to-date news, today's market data, or the latest documentation of newly updated packages, call:
     `{ "web_search": true }`
3. **Image Generation (DALL-E 3)**:
   - When the user requests an image, logo, or icon design, request it via `chatgpt_ask` and present the returned `imageUrls` directly to the user.
4. **Heavy Document & File Analysis**:
   - When analyzing CSV, Excel spreadsheets, PDFs, or large source code files, attach them via `file_paths` or `image_paths`.
5. **Continuous Conversations**:
   - Always track and pass `conversation_id` to continue in the same thread, or pass `new_chat: true` when starting an unrelated topic.
6. **Finding Past Chats**:
   - Call `chatgpt_list_conversations` to search for existing topic IDs before resuming a specific past conversation.

📄 License

MIT