#Codex is really well-tuned for my

1 messages · Page 1 of 1 (latest)

radiant oracle
#

thread

#

From @hollow drum

if you really want to use GPT and codex in that way without API, then you'll need to make one of those developer "apps" which is just an MCP server. That will let you connect GPT to that, and then that to your local stack. For Codex, it already has local access.

I'm fine with creating any tooling, API, MCP, ChatGPT application, Codex plugin, etc.

native junco
#

My first least path of resistance is that you could probably utilize the computer use feature in Codex, and tell it under what conditions to use the ChatGPT app on your machine to converse with it, I would think it should be able to type and read replies

radiant oracle
#

Yeah, but that's clunky, I'm looking for a real interface.

hollow drum
#

The only thing with computer use, is it can quickly eat up usage depending on how long and what you're doing.
The conversation interface in GPT is the hardest part to incorporate into systems using Oauth, just because it needs that web connected endpoint for the app.

native junco
radiant oracle
#

I could decide to extend it later, like to get Claude and Gemma to chat about some topic raised by GPT. I was thinking about async notes in GH Issues - kinda formal but effective. But a Discord or Slack channel seem much cooler.

native junco
#

Actually I dont know if you can use the cli with acpx hang on let me look

hollow drum
#

If you have claude, that's a completely different story. The claude desktop app can do something like this, but include GPT in the conversations.
The cowork feature has Dispatch, dispatch can use almost all of the tools available, so it can connect to the chrome extension of claude, which can connect with browser sessions, including GPT

radiant oracle
#

Oh, here's part of the challenge - to get "ChatGPT" to talk to anything it needs the context from that "application/UI", otherwise it's just another instance of a model. So one of the interfaces from the ChatGPT product is required for this to be what I've described. Otherwise I can just use a sub-agent from Codex. (Which I might do soon, just to do this, and try to load it with my personalizations.)

native junco
#

Well, really, you can use the chatgpt-cli alone, and just create a skill to use it

#

thats what I would do tbh

radiant oracle
#

(looking at the kardolus repo)

native junco
#

Just use the skill creator to create the skill, just be detailed on the intent of the functionlity you want

#

'Under X Y Z conditions consult with ChatGPT using chatgpt-cli (https://github.com/kardolus/chatgpt-cli) please build a skill to accurately utilize the cli commands blah blah blah'

radiant oracle
#

I'll do my homework but is the "ChatGPT" CLI just invoking the model or is it somehow linking into the ChatGPT data for instructions and context? I don't think that's possible which is what makes this a different kind of challenge.

#

And I'll second-guess the premise. I'm assuming Codex 5.4 is somehow different from ChatGPT 5.4 (I know they are but maybe not too much so) ... and that's why I'm trying to get the two together. If I can invoke better out-of-the-box responses from Codex or just integrate with a sub-agent 5.4 with a lot of docs for context (what Cancrit suggests) maybe this notion of integrateion with ChatGPT is just a red herring.

native junco
#

Hmm you're right ill be honest, I jumped the gun cause it had 1k stars and seemed feature packed, but it actually has no authentication

#

I have another idea

#

cause now im curious myself lol

radiant oracle
#

On phone, back in less than an hour. Thanks for the discussion!

native junco
#

It legit does not seem like there is an actual API for interacting with your official chatgpt account with memories etc. the only thing I can think is my original suggestion while may feel clunky might be the only option for now.

radiant oracle
#

back

#

I hear ya and that's why I decided to ask our cohorts here. I understand the company really doesn't want any interfaces to the front-facing platform - could be abused. I think the reasonable tech move would be to figure out why Codex isn't as broad thinking, or just give a secondary agent the info it needs to do what I want - which is not bad at all.

As I think about this I can use the Codex SDK and the Agents SDK and it should all work.

Will bang something out and post back here, whenever.

TYVM!