ChatGPT Web Relay for Hermes

This skill provides a Playwright-based browser automation that connects Hermes to the ChatGPT web interface (chatgpt.com) using your existing logged-in Chrome session. It replicates the functionality of the "GPT Relay" Codex plugin but for Hermes Agent.

Why Use This?

• No API quota consumption — uses your ChatGPT Plus/Pro/Team subscription via web UI
• Access to web-only features — GPT-5.5 Pro Extended, Deep Research, native image generation, file upload/analysis
• Persistent login — reuses a Chrome user data directory, stays logged in across sessions
• Full Hermes integration — call as a tool from any Hermes task

Installation

# Install Playwright and Chromium (if not already done)
pip3 install playwright playwright-stealth
python3 -m playwright install chromium

# Verify
python3 -c "from playwright.sync_api import sync_playwright; print('OK')"

Configuration

Create a config file at ~/.hermes/chatgpt_web_relay.yaml:

# Chrome user data directory (persistent profile)
# Use a dedicated directory, NOT your main Chrome profile
user_data_dir: "~/.hermes/browser_profiles/chatgpt_relay"

# Chrome executable (optional, auto-detected)
# channel: "chrome"  # Use real Chrome instead of bundled Chromium (bypasses Turnstile)

# Default model/mode settings
default_model: "gpt-5.5-pro-extended"  # or "gpt-5.4-thinking-light", "deep-research", "gpt-4o", etc.
default_effort: "high"  # for thinking models: "low", "medium", "high"

# Timeouts (ms)
navigation_timeout: 60000
response_timeout: 300000  # 5 min for Deep Research

# Headless mode (false = headed, more reliable for Cloudflare/Turnstile)
headless: false

# Stealth settings
stealth: true
disable_blink_features: true

Usage from Hermes

As a Python function (in execute_code or skill scripts)

from chatgpt_web_relay import ChatGPTWebRelay

relay = ChatGPTWebRelay(config_path="~/.hermes/chatgpt_web_relay.yaml")

# Simple prompt with default model
result = relay.ask("分析這個 Python 代碼的性能瓶頸", model="gpt-5.5-pro-extended")

# Deep Research
result = relay.deep_research("研究 2024 年台灣半導體產業趨勢")

# With image upload
result = relay.ask_with_image("分析這張圖表", image_path="/path/to/chart.png")

# With file upload
result = relay.ask_with_file("總結這份報告", file_path="/path/to/report.pdf")

# Switch model mid-session
relay.switch_model("gpt-5.4-thinking-light", effort="high")
result = relay.ask("用深度思考模式解決這個數學問題")

print(result["text"])
print(result["images"])  # generated image URLs/paths
print(result["conversation_url"])  # link to chatgpt.com conversation

As a Hermes tool (via delegate_task or skill)

# In a Hermes skill or delegate_task context
from hermes_tools import terminal

# Quick one-off via terminal
script = '''
from chatgpt_web_relay import ChatGPTWebRelay
relay = ChatGPTWebRelay()
result = relay.ask("用 GPT-5.5 Pro Extended 分析: {{prompt}}")
import json; print(json.dumps(result, ensure_ascii=False))
'''
# ... execute via terminal

Core Functions

Supported Models/Modes (web UI names)

Architecture

Hermes (Python) 
    │
    ▼
ChatGPTWebRelay class (this skill)
    │
    ▼
Playwright → Chrome (persistent profile, user_data_dir)
    │
    ▼
chatgpt.com (logged-in session)
    │
    ├── Send prompt → wait for response streaming to complete
    ├── Handle model switcher UI
    ├── Handle file/upload dialogs
    ├── Handle Deep Research progress polling
    └── Extract: text, images, conversation URL
    │
    ▼
Return structured dict to Hermes

Selectors & UI Automation (ChatGPT Web UI as of 2026)

The skill uses robust selectors with fallbacks:

• Prompt textarea: textarea#prompt-textarea, textarea[data-id="root"], div.ProseMirror[contenteditable="true"]
• Send button: button[data-testid="send-button"], button:has-text("Send")
• Model switcher: button[data-testid="model-switcher"], button:has-text("GPT")
• Model options: div[role="option"]:has-text("{{model_name}}")
• Response container: div[data-message-author-role="assistant"], .markdown.prose
• Image results: img[data-testid="image-generation-result"], a[href*="files.oaiusercontent.com"]
• Deep Research progress: div:has-text("Researching"), div:has-text("Analyzing")
• File upload: input[type="file"], button with data-testid="attach-files"

Pitfalls & Solutions

Example: Full Hermes Integration

# In a Hermes skill or cron job
from chatgpt_web_relay import ChatGPTWebRelay

def hermess_chatgpt_relay(prompt: str, task_type: str = "analysis") -> dict:
    """Hermes-callable wrapper."""
    relay = ChatGPTWebRelay()
    
    try:
        if task_type == "deep_research":
            result = relay.deep_research(prompt)
        elif task_type == "image_analysis":
            # expects prompt to contain image path or be passed separately
            result = relay.ask_with_image(prompt, image_path=...)
        else:
            # Default: use strongest model
            result = relay.ask(prompt, model="gpt-5.5-pro-extended", effort="high")
        
        return {
            "success": True,
            "response": result["text"],
            "images": result.get("images", []),
            "conversation_url": result.get("conversation_url"),
            "model_used": result.get("model_used"),
        }
    finally:
        relay.close()

Security Notes

• Never use your main Chrome profile — create a dedicated user_data_dir
• No credentials stored — relies on existing browser login state
• Runs locally — no data leaves your machine except to chatgpt.com
• Rate limits — respect ChatGPT web UI limits (not API limits)

Files in This Skill

chatgpt-web-relay/
├── SKILL.md           # This file
├── chatgpt_web_relay.py   # Main Python module
├── config.yaml.example    # Example configuration
└── scripts/
    └── hermess_wrapper.py # Hermes integration helper