Three Ways to Call Codex from Claude Code — A Practical Breakdown
This article explores three methods for calling OpenAI Codex from within a Claude Code session, including their characteristics and when to use each approach.
Why it matters
This article provides a practical breakdown of the different approaches for integrating Codex into Claude Code, helping developers choose the most suitable method based on their specific needs and requirements.
Key Points
- 1The codex-plugin-cc uses the App Server Protocol (ASP), not the raw CLI
- 2Method 1 (codex --full-auto) is a one-shot CLI process, while Methods 2 and 3 use the ASP mode with a resident broker
- 3Method 2 (codex-companion.mjs) provides job tracking and background execution, while Method 3 (codex:rescue sub-agent) offers prompt optimization
- 4All methods support both ChatGPT subscription and OpenAI API key, but subscription-only features like Fast Mode are unavailable with API keys
Details
The article provides a detailed comparison of the three methods for calling Codex from Claude Code. Method 1 uses the raw CLI with the `--full-auto` flag, which is a one-shot process that interacts with Codex via stdin/stdout. Methods 2 and 3 leverage the codex-plugin-cc, which internally uses the App Server Protocol (ASP) and a resident broker process, allowing for features like job tracking, background execution, and thread continuation. Method 3, the codex:rescue sub-agent, also offers automatic prompt optimization using the gpt-5.4-prompting skill. All methods support both ChatGPT subscription and OpenAI API key authentication, but subscription-only features like Fast Mode are unavailable when using an API key.
No comments yet
Be the first to comment