Split work between ChatGPT architect and Codex executor

The poster built a bridge where ChatGPT plans and reviews while Codex does the local execution work.

What They Did

The poster used MCP-style local file access to connect ChatGPT with a local Codex setup, after seeing ChatGPT handle a local CAD task by asking clarifying questions and working with subagents for nearly an hour. They then built a small workflow layer that assigns ChatGPT the planning and review role, while Codex handles local execution steps like touching files, running commands, and checking results, with basic safety boundaries added by the poster.

How It Went

The immediate CAD test showed stronger high-level capability from ChatGPT than Codex, though the CAD output itself was only so-so. The bridge itself is presented as experimental, with no performance numbers yet and the author asking others how they would design the workflow.

Worth Knowing

It was built for the Windows version of Codex, and the author said using computer-use directly inside native apps like the ChatGPT desktop app usually failed, so they bypassed that with Codex web Chrome control.