Codex CLI Keeps Getting Better (Everything New)

Okay, so this is everything new that was added  to Codex CLI over the last week. First of all,   you want to make sure that you have Codex CLI  updated by running this command over here,   and you should be on version 0. 44. Anyways,  we have a brand new UI. So if you go to Codex,

then you can see it looks a tiny bit different  because there's a new design right at the bottom   over here. I don't think it looks quite as good  as the Claude Code design, but it certainly looks   better than the previous design because you can  see right over here, like the input box, it just   didn't look particularly nice. It seems they also  removed this thing at the bottom that shows the   shortcuts that are available. And replaced it  with this like question mark. So you can press   shift question mark and then see those same  shortcuts over here. They also improve the   model picker as well. So if you do slash model,  then you can see that you have two choices over   here. And then when you select one of them, you  can choose a reasoning effort. Previously it was   you had the reasoning effort and the model name  together. So for example, going back over here,   this is what it looks like before. It was much  more messy, but now it looks slightly nicer.

Custom prompts now support named and positional  arguments. So to set up a custom prompt,   you can go to your Codex folder by doing CD and  then like wavy line slash dot codex and then press   enter. And that should be in your codex folder.   And then you can make a new directory by running   this command, press enter and then CD inside of  it. And now you can make a new custom prompt. So   I'm going to call it like super reviewer. And it  does have to be a dot MD file, a markdown file,   press enter. And then I can say, can you do a  comprehensive review of $1? And that $1 question   mark will be like the positional argument  that's involved. So then I can press exit,   save this, enter. And if I reload codex  by exiting out of it, doing codex again,   then I can do slash. And then you can see I have  this custom prompt super reviewer, press tab,   and then I can like mention the file and say like  index. ts, press enter. Then if I do control T,   you can see right over here, the user message is  can you do a comprehensive review of index. ts?    So that was a positional argument it accepted.   I can also add in more positional arguments by   doing like and $2, $3, $4, and $5, and that would  be like positional arguments. Can you review the   git commit, $commit in capital letters? And then  it appears as this, so I can press save. And if I   close Codex and then run it again, then I can  trigger super reviewer. And when I press tab,   you can see automatically gives me like the  key value that I can fill in. And let me just   make up something over here because like I'm not  actually sure what the commit names are. And now   you can see it's trying to find that commit, which  I can't find because I just made it up. And yeah,   this can be useful for any prompts that you use  often and are sick of typing again over and over.

They also now added support for HTTP streamable  MCP servers. So previously when it came to using   MCP servers, for example, the Exa MCP, you'd  have to define it kind of like this and use the   NPX command. And you can see it's working here  on the left hand side. But now I can completely   remove this and then just write in URL equals and  then copy and paste the Exa MCP URL over here and   then add bearer token equals and then I can run  codex MCP list and you can see it appears in the   list of MCP servers. And it can also run Codex,  MCP get exa, and you can see that it supports   streamable HTTP. But it seems that now for some  reason when you run Codex and then run slash MCP,   it doesn't actually appear on the list despite it  appearing when you do Codex MCP list. So it's like   there's a discrepancy over here. I can see that  someone opened up an issue about this on GitHub,   so it should probably be fixed by the time  you're watching this video or something like

that. Another pretty neat feature they seem to  have added is Codex Cloud. So currently on the   ChatGPT version, When you go to chatgpt. com/codex,  then you can see all the different cloud versions   that you ran with codex running in the cloud. So  for example, in this project over here, I have   fixed library missing error in llama-server,  and I have these different versions I had run,   and no changes were made because I didn't press  start task. It just gave a summary of the changes   that would have been made. But now if I go to  warp and then I run codex cloud, press enter,   then I can see the same task also appears on  here. I can press enter, and then I can press   tab to cycle through their different attempts. As  it shows here. And then I can press A to apply any   changes that have been made. But it says no diff  available to apply because it didn't actually make   changes since I didn't approve the task. I think  if I press enter, then it should open any changes   that were actually made, but no changes were made.   I can also press new task by pressing N and then   give it a task like find a bug and fix it. And I  can press control N to cycle between the number   of attempts that will happen in parallel, which  is exactly like the web version where you have   version 1, 2, 3, and 4. So I'll press down, have  three of these running in parallel. Press enter,   and I can also press ctrl+o to change the  environment that it's running on. But I only have   one environment right now. Press escape, press  enter, and that will set off the task. And then   I can see the task is pending over here. Press  refresh to reload it as well, and press enter,   and I can see the conversation so far. And then  finally I can press quit as well. And ultimately   I think that this is slightly nicer because I knew  that ChatGPT, the web version of codex, did exist,   but I rarely used it myself because I didn't like  going on the website and switching between the   two. But now it seems that you can do the cloud  version all in your terminal, which seems pretty   handy. So you can probably do something like if  you're on a train and you have like pretty spotty   internet, then you can set up a cloud version  and then review the cloud version once you're   done because it will be stable in the cloud.   Another useful thing is that it now shows the

context window usage whilst tasks are running. So  if you give it a task such as give me a summary of   the project, press enter, and then you write slash  status, then you can see it shows the exact amount   of tokens that have been used. And then if I run  it again slash status, you can see the number has   now been updated as well. So I think that can  satisfy any curiosity that you may have because   you can see the exact amount of tokens that have  been used instead of just looking at the general   percentage left instead. And there are tons more  small quality of life improvements. So yeah,   I think that Codex CLI is getting really good. I  think in a couple months it will have caught up to   Claude Code in terms of features. And yeah, I just  really enjoy using it. I was actually going back

and forth between Claude Code on the left hand  side and Codex CLI on the right hand side, and   then getting Claude Code to do the implementation  and Codex CLI to check the implementation that   Claude Code was actually doing, and then giving  back the tasks to Claude Code to then complete.    And I did make a separate video about it. And  it seems that Claude code with Sonnet 4. 5,   whilst it does do a better job at completing the  tasks, it still just ignores the fact that there   are still some outstanding tasks. So you can see  over here, there are still some outstanding tasks   that Codex CLI identified based on this like fix  that I was making to my application. Hyperwhisper   that you can see here, there's a coupon code  down below if you are interested. And despite   some back and forth, it kept failing to actually  execute and finalize some of the last remaining   tasks. And then I said to Codex CLI, can you just  fix it instead? Because Claude Code kept failing   to fix these particular issues. And yeah, it does  a pretty good job, as you can see over here. So,   so far my workflow has still been using Codex  CLI and Claude Code together in parallel, but   I have been experimenting with a bunch of other  models as well because I'm genuinely interested   in the capabilities of the latest models. So I'm  like trying a bit of everything really. But my go   to workflow is still using Claude Code and Codex  CLI together, relying more heavily on Codex CLI   to do like high level tasks and relying on Claude  Code to do like the lower level implementation.    And then in some cases just relying on Codex CLI  in isolation. Because Claude Code still fails on   certain tasks such as this particular one. And  yeah, if you are interested in this application   that I was working on, then it should be linked  down below with a coupon code as well. And if   you do want to learn how to make profitable  applications yourself, like the one I showed in   this video, then I do cover a lot to do with that  in my AI Startup School that you can join using   the link in the description down below. There are  a bunch of classes that teach you everything about   vibe coding, vibe marketing, coming up with ideas,  distributing, a bunch of case studies, and really   comprehensive classes, all about Codex CLI, and  Claude Code with some advanced techniques as well.    And a bunch of people have already joined as you  can see here and have had good success with their   own applications as well. And one of the biggest  benefits that you'll get from joining is that   you can ask me any question that you have and get  support from me for your own applications as well.