Choose the orchestrator agent from providers other than OpenAI

[5ocworkshop]

What feature would you like to see?

I know this was discussed a while back as something that was being considered, is it still on the radar to allow other models to be the orchestrator agent?

I recently switched from OpenAI $200/mo to Claude to try it out after a year of being away from it. But I'd really like to keep using just-every Code, it's fanastic what you've done with it. I know I can use Claude as a sub-agent but I don't see any way to use it other than that?

[5ocworkshop]

I haven't had a chance to try it yet, but last night I got Claude to add itself in, in a hacky kind of way, to the tool. It's not fully integrated, but it's a start. :)

[cbusillo]

Neat! I would love to know how this works for you. In my experience Opus burns up tokens/limits so much faster. What is your experience?

[5ocworkshop]

I only just switched to it, but I'm actually seeing the opposite but I think that has a lot to do with my workflow which I've been refining for a long time. Compared to the Codex subscription I'm getting more done with my 5h allotment. However, I see a LOT of tokens being wasted running basic bash commands and read irrelevant output so it really has me thinking about wasteful that is and how maybe it could be reduced.

I was going to say "Maybe it's the application chewing up the tokens?" but I think you mean using Opus in a different context since you can't make it the orchestrator today? Or do you mean calling Opus from within 'code'?

[cbusillo]

I meant, in my experience, Opus uses its rate limit up quickly in Claude code.

It's great to hear it's working for you with Code. Maybe it will become a general feature for us to try.

[zemaj]

This is interesting! Do you have a branch/fork I could take a look at?

[freelanceagent1]

This is interesting! Do you have a branch/fork I could take a look at?

This is interesting! Do you have a branch/fork I could take a look at?

Sir, https://github.com/freelanceagent1/gemini-cli-proxy I did a translation proxy for myself, forked and updated from a different repo. It still sometimes stutters in /auto mode, can't pinpoint why for now (my guesses are either I did not fix all the translations properly or the way codex and inherently code ping the model provider and get rate limited by google), but maybe this will give you a broader idea on how to potentially integrate similar functionality into Code, love the tool.