OpenAI Codex engineering lead Thibault Sottiaux says Codex tools can now be configured to run any open-source model

How’s glm in the codex harness

@thsottiaux Hey Tibo, is it possible to set up a hybrid approach in Codex where we use GPT models as architects and open source models as executors?

I'm trying to get it working, but it seems Codex expects the same tool-calling protocol from other models, which they don’t have.

TIL you can use open-source models in codex

@FilipBaturan @thsottiaux Yes! You can edit Codex CLI and build whatever routing layer you want.

I use a Custom Pi setup called Takomi: GPT orchestrates, other models execute. It’s my hybrid architect/worker flow. Happy to keep improving it with others: https://github.com/J-StaR-Films-Studios/VibeCode-Protocol-Suite

@thsottiaux Can I use a GPT model to orchestrate other models as subagents?

@thsottiaux If you want to use third-party models in Chat Completions format with Codex(which is very common like DeepSeek), feel free to give CC Switch a try😜 https://www.ccswitch.io/en/tutorials/codex-deepseek-routing-guide

@thsottiaux And also we can use chatgpt pro for planning and codex or any other harness for implementation or go full rampage with chatgpt

@thsottiaux

@bxsck_ @thsottiaux Yes you can, I've done this many times. Once you add model provider and model/name, you can use it in agent toml file. Ask Codex to do it for you.

@FilipBaturan @thsottiaux You should try this , i built it exactly for this workflow

@thsottiaux @thsottiaux please fix windows codex app

@thsottiaux Wow, how did I not know this. Thanks Tibo

@FilipBaturan @thsottiaux Yes sure, just set up codex app itself with the openai frontier model, then tell it to spawn codex cli with the models of your choice for work. In fact, just put this post and yours into codex now and tell it to write a good AGENTS.md that will do this, job done

@thsottiaux It can be configured globally, but it can’t be mixed. In one case, I want to use GPT-5.5 as the main intelligence model, while using another provider for subagents. In another case, I want to use 2 different providers, neither of which is OpenAI. not enough flexibility.

@FilipBaturan @thsottiaux not so well and efficient as it would be with say @opencode . just use codex as the architect, write a tasks.md file and then execute with opencode and any agent you like

@thsottiaux opensource codex please

@thsottiaux The model picker on desktop doesn’t show external provider models atm due to a bug.

I did exactly this yesterday. I tried to make the main agent orchestrate subagents with external models.

And it worked, but you need a proxy that handles the non-standard tools Codex is providing.

Wasn't going to make it public yet, but in case it helps, here's the trick I'm testing so far:

https://github.com/antoniolg/codex-provider-shim

@trashpandaemoji @thsottiaux https://github.com/0xSero/codex-shim

@thsottiaux 🫶 that's actually very cool - let's do even more