adamallcock/codex-chatgpt-control

What it solves

codex-chatgpt-control provides a structured bridge for AI agents to interact with the visible ChatGPT web interface (chatgpt.com). It solves the problem of users having to manually move work between local execution environments (like Codex) and the ChatGPT web product for tasks like conversational review, research, or creating deliverables.

How it works

The SDK acts as a runtime authority that drives the browser via a compatible Codex/browser bridge. It does not use private APIs or hidden endpoints; instead, it interacts with the user-visible UI controls of the signed-in ChatGPT session. It can detect whether the user is in the "Chat" or "Work" experience, inspect and apply configuration settings (such as model or effort levels), and manage threads and tasks. The project provides both a Node.js runtime and a Python parity client that communicates with a Node backend.

Who it’s for

Developers building agents that need to delegate specific workflows to the actual ChatGPT web product while keeping the process visible and user-directed.

Highlights

• Visible-Session Only: Operates exclusively through the browser UI to ensure transparency and user control.
• Experience Detection: Automatically distinguishes between ChatGPT Chat and Work experiences and manages the corresponding controls.
• Work Task Management: Supports starting, polling, steering, and reading results from long-running Work tasks to prevent duplicate submissions.
• Artifact Handling: Capable of uploading local files and downloading generated artifacts, including image-only results.
• Agent-Facing Skills: Includes a Codex plugin and skills that provide operating guides for agents to use the SDK consistently.
• Privacy-Preserving Reports: Generates local run reports that omit prompt and response content by default.