I built a bridge between ChatGPT and Codex
这条记录涉及编程工具或代码能力更新，适合开发者评估工作流变化和可复用价值。
作者发现 ChatGPT 具备类似 Codex 的能力，其背后是一个强大的多智能体系统。通过让 ChatGPT 直接读写本地文件，作者成功让它完成了一个本地 CAD 任务。在任务中，ChatGPT 像 Claude Code 一样逐一询问需求，并利用多个子代理工作近一小时才完成。尽管 CAD 结果一般，但作者认为这预示着 OpenAI 的下一步发展方向。…
Recently I found that ChatGPT actually has Codex-like capabilities as well. Behind the scenes, it is essentially running a powerful multi-agent system and has many skills.
There has already been work using MCP to let ChatGPT directly read and write files on a local PC. I used this approach to let it complete a local CAD task for me. Surprisingly, its capability was far beyond Codex. Like Claude Code, ChatGPT asked me about my specific requirements one by one, and worked for nearly an hour using multiple subagents to complete the task. Although the CAD result itself was just so-so, this task gave me a glimpse of OpenAI’s bigger next move.
That made me think there may be a bigger direction here. ChatGPT seems better at high-level reasoning, planning, and reviewing, while local coding agents like Codex are better positioned to execute small steps, touch files, run commands, and verify results.
For some tasks, it feels more practical to let ChatGPT act as the architect/reviewer and keep the local agent focused on execution. So I built a small agentic skill to standardize this workflow and add some basic safety boundaries.