Breakeven Brothers

[00:00] Speaker: We’re live, and as we’re live, we’re trying to remember what episode we’re on. So we’re just winging it this morning. But Brad, how’s everything going?

[00:15] Brad: It’s great. It’s Friday. It’s May 1st. Ready for the weekend. So yeah, how about you?

[00:21] Ben: It’s good. It’s good. Been a busy week. You were in town with your lovely wife and our parents last week, so that was nice to see everybody.

[00:28] Brad: Yeah, had a good time. We had a gauntlet. Remember the gauntlet? We played probably over 150 games of Soulcalibur.

[00:39] Brad: So, if you have a Nintendo Switch, the new one ships with these old retro games, and Ben and I played the original Soulcalibur. So we were playing that, we got into it, and then we went off the deep end and played many, many series of Soulcalibur, which was a lot of fun.

[00:59] Ben: Yeah, it is episode 41, by the way. That’s what we’re on.

[01:03] Ben: And yeah, Soulcalibur, it was very intense. Brad bested me, unfortunately, but it was good.

[01:13] Brad: Voldo’s too good out there. If you know who that is, go look him up if you played those games. I think once you see a photo, you’ll be like, “Oh, I remember that character.”

[01:22] Ben: Infuriating. He’s an infuriating player to play against.

[01:24] Ben: Yeah, geez. Well, cool. So what’s shaking in the AI world? What should we kind of catch up on? I think we wanted to catch up on some cool Codex stuff, which is great because I have my own questions as a Codex newbie.

[01:36] Ben: And just for listeners, when Brad was with us in town, he was working on stuff and doing stuff, and the entire time I saw him on his computer, he was on Codex. So it’s a testament to, like, this thing is being used through and through by the most experienced and knowledgeable developer that I know.

[01:54] Ben: So lay it on me, Brad. What’s new with Codex, or what should we catch up on? What do I need to know about it?

[02:00] Brad: Yeah, that’s a great intro. I think before that, I would like to just quickly highlight GPT-5.5 because I think this model in particular is just an absolute powerhouse.

[02:10] Brad: And the reason it’s important for Codex is because Codex is just so much better with a better model. And I know we had talked last episode about bringing back MCPs. I think this whole picture, when I think about Codex, is just this amazing harness.

[02:28] Brad: And for folks listening, I’m using the Codex app. So I use Mac. There’s a downloadable Codex Mac app. That’s what I use. You can also use a CLI, but I’ll talk about a few reasons why I use the app.

[02:40] Brad: But yeah, Codex is my daily driver. This is something that I use for coding, so adding new features. This is something I use for data analysis, so writing SQL queries, pulling data, connecting with systems. Also to do web search at times.

[02:56] Brad: There are often times I’m trying to think about the right approach to things, and I have a little bit of code, I have a little bit of an idea, and I just want to get better grounding. So there becomes a world in which we have ChatGPT, an excellent AI communication tool that just helps you get your job done, and then Codex is popping up as this very, very effective tool where sometimes it’s hard to figure out which one to go to.

[03:18] Brad: And quite honestly, I’ve been leaning on Codex a lot more. So yeah, I’d be happy to talk about some of these things.

[03:24] Brad: But I think overall, GPT-5.5 is out in ChatGPT, it’s out in Codex, it’s even out in the API. And really, the intelligence is excellent. It feels like it can unblock itself, is how I’d put it.

[03:38] Brad: So there are lots of hurdles in getting things done. It feels like GPT-5.5 in Codex can just get past things that it wouldn’t necessarily do before.

[03:48] Brad: So if I ask it to do something complicated, maybe it goes down 20 steps, and at one of those steps it fails. GPT-5.5 will figure out a workaround, go Google something, go look at different source code. It’s pretty incredible what it does.

[04:01] Brad: And the big unlock for me is that I’m able to give it a harder task, and bigger tasks, because it can unblock itself.

[04:09] Brad: So I think with the advent of GPT-5.5 and the better Codex app experience, I’m doing more things. I’m doing more complicated things, and I’m doing that, ideally, all at once. That’s kind of the baseline.

[04:22] Brad: And then when we talk about the Codex app, there’s a lot to it. There’s automations, there’s plugins, there’s a review pane, all this stuff we’ll dive into.

[04:32] Brad: But I think at a high level, if you’re not using the Codex app yet, hopefully by the end of this session, you’ll be convinced on why to use it and to give it a try.

[04:41] Brad: And honestly, it evolves pretty often, so if there’s something that you want to see, feel free to reach out in the comments and we’ll see what we can do. I think it’s a very evolving surface, and developers love it.

[04:54] Brad: I think it’s starting to edge into the knowledge work arena, so Claude Co-Work. But honestly, it’s evolving so much, and it’s only gotten better.

[05:02] Brad: So from where we’re at on May 1st, if we take a step back in two months, I imagine it’s going to be 10 times better, because the entire time that I’ve been working with it, honestly, it’s gotten better and better. So I’ll pause there. Any questions, Ben, before I jump in?

[05:16] Ben: No questions, but just more observations, because I think you and people kind of in the engineering space on Twitter are always the leading indicators of where things kind of go.

[05:27] Ben: And then I have my accounting world and my finance world—that’s my background and where I work. And it’s funny because probably back in last year, May or June, you were pushing Claude Code, saying the terminal is the best place to be. And I was like, “I don’t like that. I like my IDEs. I like Cursor. I like Visual Studio.”

[05:50] Ben: Then eventually I came around, and I’d say—I don’t know when exactly I kind of got Claude Code going—but probably late in the year of 2025. And so I’m on Claude Code. I’m enjoying it. I’m liking the terminal. I also use the Codex CLI.

[06:01] Ben: But then Brad goes, “Oh no, no. Now we’re doing the app. It’s not about the terminal anymore. We’re doing the app.” And so I’m kind of catching up to that.

[06:10] Ben: So I have Codex downloaded on my personal computer, and I’m trying to kind of start to use it, but I’m still pretty used to the terminal now. And I use the Codex CLI as my daily driver, but I don’t use the app quite yet as the daily driver.

[06:26] Ben: But it’s interesting too, because similar to me kind of lagging in terms of what all the engineers and people in your space are adopting, accounting is similar. I see people kind of just getting into Claude Co-Work more—not necessarily Claude Code, but everyone’s kind of going into Claude Co-Work.

[06:46] Ben: And I think for me, it’s like if you like Co-Work, but maybe you already have an OpenAI subscription, you can get Codex. It’s not the exact same, but it’s a similar application where you can do all the cool stuff with your local files on your computer, but it’s with your OpenAI subscription that you pay for.

[07:08] Ben: So I definitely want to keep using it. I have used 5.5 extra high reasoning, and it’s pretty solid. Yeah, I’m like, “This is pretty sweet.”

[07:19] Ben: So one of the use cases I did set up and was just tinkering around with was I hooked up the Canva MCP. So on the Codex app, I have it connected to Canva.

[07:30] Ben: And then from there, being able to prompt it to make designs based on a style guide that I have in a project folder. So in a project folder, I just said, “Hey, these are the fonts I want. These are the kind of main color themes that I have.”

[07:44] Ben: And then use the Canva MCP to go create a brochure of how to reconcile cash accounts, stuff like that—accounting stuff. And it does a pretty good job.

[07:53] Ben: But I think some of my issues with it are on the Canva side, because basically the only way that you can interact with the MCP is through their prompt. So you kind of prompt it, and then the Canva side does like a magic AI image-gen type deal.

[08:11] Ben: And that’s not really part of Codex’s thing. Codex just sends the prompt and then gets the response. But it’s pretty cool, and as someone who does not like marketing at all and has no creative bone in their body, it’s starting to feel like you can unlock some cool stuff with that.

[08:25] Ben: And yeah, it’s easy to set up within Codex. So I was really surprised how easy that was, given that MCPs in the past have been kind of tricky to set up just from the very start. And they’ve gotten easier and easier, but now it’s like two clicks and you’re connected.

[08:46] Brad: Yeah. When we talked about CLIs back then, it felt like that was the entrance into high intelligence, and so I leaned toward CLIs because Claude Code was so good when it came out.

[08:55] Brad: And over time, it got a bit bumpy due to so much demand. I think if they just had five customers, it would be amazing. But since the demand was so high, people loved it, and it got a little bit shaky.

[09:00] Brad: And once the UI apps came out—I think Claude Co-Work came out at the end of last year. Honestly, AI timelines are hard. When that came out, I wasn’t super excited about it, but I tried it and I thought, “Oh, this is actually kind of nice.”

[09:10] Brad: I’d been so in the terminal because I had to be, and being outside of the terminal, I thought, “Oh, it’s not that bad.” I think a lot of people felt that to be more approachable.

[09:18] Brad: I think the app, in essence, was to bring better and easier features. For example, in Codex, you can have plugins. So we talked about MCPs—that’s kind of the underpinning foundation of all these things—but plugins are an easy way to install behavior that almost emulates an MCP or almost wraps an MCP.

[09:40] Brad: That allows you to click and connect to many services. Let me take a look at the plugin list—like Google Calendar, Outlook, etc. You click the plus button. So you go to the Plugins tab within the Codex app, you click the plus button, and that opens up a modal to connect to these official services.

[09:56] Brad: You don’t need to fight with any MCPs. You click connect, it brings you to a login screen from that service, and you’re done.

[10:02] Brad: So I think MCPs are still available within Codex, but plugins are kind of this easier way to approach it.

[10:09] Brad: And we’ve talked about MCPs for a long time. I think they’re really, really good now, given that the model knows how to use them and it’s intelligent. But that’s one thing that you get out of the box when using some of these app-focused approaches versus CLI.

[10:23] Brad: There isn’t really an easy way, from a consumer point of view, to just get connected to Slack, get connected to Gmail. And in plugins, it’s super, super easy.

[10:29] Brad: And I think one callout that I completely forgot about is we have computer use available in Codex now—a really incredible technology that allows it to look at your Mac applications and gives Codex control to say, “Go tap on this button.”

[10:47] Brad: So for example, automating my browser—if I have to work between multiple tabs and we don’t have plugins or MCPs—computer use can do that.

[10:54] Brad: If you’re working between a spreadsheet in Excel and some other service in your browser, it can copy data from your spreadsheet and bring it over. It’s really incredible, and it’s a little bit new, so there are some rough edges.

[11:07] Brad: But if you step back and think, “This is AI; it’s only going to get better,” if today it can control my computer at a decent pace—maybe not perfect, and maybe not get everything—in three to six months, who knows where it’s going to be.

[11:19] Brad: So that is also a plugin in the Codex app.

[11:22] Brad: Outside of that, I think the Codex app brought me very front and center to worktrees. So when we talk about having a smart model—having a model that’s plugged in and has MCP servers and plugins that can go fetch data—we want to do more all at once.

[11:38] Brad: Part of this is using Git worktrees. So previously, when using Claude Code, I would work on one thing at a time. I would basically finish that one task and move on to another, and wouldn’t have many parallel things running.

[11:51] Brad: Git worktrees I’d always been confused about as an engineer—like, how does this work? I’ve heard about it. It’s a lot of effort to get this kind of orchestration set up. But with Codex, it’s basically a first-class feature.

[12:03] Brad: You can click on a button when you’re starting a new Codex chat to say, “Do I want to work locally or in a new worktree?” And that’s it. You don’t have to think about what happens under the hood. The new worktree does all the fancy Git stuff to isolate your code changes, and you don’t have to worry about it.

[12:18] Brad: It has first-class integration review tooling as well. So if you need to review changes that Codex made, that is scoped to that worktree and not to any other changes.

[12:28] Brad: So I think part of the Codex app experience is having these first-class features that you could definitely do in the CLI, but making it way, way easier to install powerful connections, work on multiple things at once, and then once it’s done writing code or doing whatever you’re doing, have a web browser built into Codex, have a file review tool built in.

[12:51] Brad: All these things are very first-class and what you would expect for both knowledge work and engineers to get done. It just feels like it’s all in one app.

[12:59] Brad: I think that part is very nice because when I did a lot of Claude Code, I was writing code in the terminal, going and reviewing code in a different app, and it just felt a little bit disjointed.

[13:10] Brad: It feels like Codex is trying to package everything into one. And again, we’re getting there. It’s not perfect yet, but with the high intelligence, easy-to-access plugins, and honestly pretty awesome automations, you can get really, really far.

[13:23] Brad: And you don’t have to be that in the know about how things work. They just work well, and I think that is part of the good experience.

[13:30] Ben: Yeah, it’s funny you mentioned the worktree stuff because that’s something that I saw and I was like, “I don’t know what this is.” And I’m in the knowledge-work space, so I’m not in engineering land.

[13:39] Ben: I know GitHub and its uses, but I’ve been using Codex for more marketing ideas and just general knowledge work, not necessarily building apps. And so those kinds of features, I’m just like, “I don’t think I want to go down that path.”

[13:55] Ben: And I’m still not comfortable enough with Codex to just try new things. I don’t want to break anything—that’s not the right word—but do something that is hard to unwind or not what I intended. So I’m still kind of tiptoeing through the app.

[14:11] Ben: But yeah, it is very nice. And I didn’t realize—just to clarify—that plugins are not MCP. Is it a different protocol under the hood, or is it still MCP under the hood?

[14:19] Brad: I think it could be MCP under the hood. I think it’s a bit of a different connection, but I’m not exactly sure how it’s packaged.

[14:26] Brad: From a user experience point of view, you click a button to connect, and you’re authenticated with that official service. It might be built off the rails of MCP—unsure—but at the end of the day, it’s a much more seamless connection experience.

[14:39] Brad: I think a lot of companies have been striving toward an automated MCP setup, and this is basically that in a different flavor.

[14:48] Ben: Yeah, no, it’s super cool. Yeah, so the Canva one I have, and then a couple other ones that I was looking at, look pretty sweet.

[14:54] Brad: Yeah, I think they’re shared between Codex and ChatGPT such that if you connect one in ChatGPT, it’ll also be connected in Codex. So there’s a nice unification there.

[15:05] Brad: But yeah, I think at a high level, when I look at Codex, you definitely spend time working with the tools. Out of the box, it’s a big surface area, and you’re equipped to do great things, but you have to learn the tool.

[15:17] Brad: Like any new tool, there’s complicated onboarding. I think part of the focus is to make that a little bit easier.

[15:24] Brad: But when I take a step back and look at Claude Code—how I used it in May of last year—and Codex now, I think it boils down to a few things.

[15:35] Brad: One, I used plan mode all the time in Claude Code. This is because it felt like the model wasn’t smart enough to understand what I was trying to do.

[15:44] Brad: So I would basically talk to my computer for three minutes, and I’m a big proponent of still doing that—dumping out as much context as I can—because at the end of the day, using AI tools is trying to be a context maintainer, a context savant, so to speak.

[15:58] Brad: Just make sure you have the amount of data that you need to get things done—but not more, not less—and that is a really hard battle.

[16:06] Brad: So I think when I was working with Claude, I would just dump a bunch of data through my voice to Claude. I would ask it to go into plan mode, or go do a bunch of research and come back with four or five questions. I’d say yes, no, yes, no. And then once it was done planning, I would say execute.

[16:21] Brad: Now it’s vastly different. I talk with my computer, but I don’t even engage plan mode. I believe there is plan mode within Codex, or at least there used to be, but I don’t use it anymore.

[16:33] Brad: And I think the raw reasoning behind why I don’t use it anymore is because these models understand me a lot more. And if I do want to get more information before I work, oftentimes I just tell it, “Don’t write any code. Let’s just talk about it.” That is basically the equivalent of plan mode.

[16:50] Brad: And back then, models weren’t as smart. When plan mode was engaged, I think there was no ability for the model to invoke writing files. Whereas if you’ve used old models in the past and you’d say, “Don’t write any code. Just chat with me about it,” it would actually start writing code.

[17:05] Brad: So I think the instruction following for these new models is really good. You don’t need this very coerced plan.

[17:20] Brad: So I do planning, and I try to orchestrate that, but it feels like you’re talking to somebody. I think that is a big unlock of, “Hey, I want to talk about changes, then let’s build it.”

[17:23] Brad: But to be quite honest, oftentimes I give it a lot of context and I say go, because it is able to follow instructions. It does have powerful tools connected to it. It has very, very smart intelligence and reasoning to unblock itself, such that at the end of the day, it mostly gets what I want.

[17:38] Brad: Not perfect, but that’s one of the big differences I see from Claude Code of May 2025 to Codex now: try not to use plan mode. Try to give it bigger things than you’d think.

[17:50] Brad: Because Claude Code back then was very easy to use for small things and do a lot of small things. I think with Codex, you need to open up the horizons and think about, if AI could do anything, what do I want it to do now? And how much information do I think I need to give it for it to be successful?

[18:06] Brad: And the closer you are to that domain, the easier it is. For me as an engineer, I still give it very detailed prompts. I don’t really go hands off the wheel and say, “Go build this feature.” I say, “Build this feature, keep this in mind.” And all of that extra context of me understanding the system as much as I can makes the quality better.

[18:22] Brad: So I oftentimes see things like, “We outsource the intelligence, but not our understanding.” And I think that part is critical, because your understanding of the system helps influence AI to make a better end product.

[18:38] Brad: You can still understand your intent and where you’re trying to go, but if you can guide it in a way that makes it better, that is still extremely valuable, no matter what domain you’re in.

[18:47] Brad: So I try to let Codex go hard at a really big task, and I try to do multiple things at once, and that’s where worktrees come in.

[18:53] Brad: And I think finally, to round out the major differences, is this whole review mode.

[19:01] Brad: So review mode is a right-side pane within Codex. That means once I ask Codex to do something, there’s the main chat window in the center. In the top-right corner of the UI, they allow you to expose this review mode, which essentially is a file browser that’s built into Codex.

[19:17] Brad: You’re able to review the lines of code that have changed. The reason this is important is you can actually comment on the lines of code that Codex has changed and ask questions.

[19:27] Brad: So you could say, “Why did you change this from X to Y?” and make a bunch of comments. These are all then fed back into the conversation. So you could say, “Hey, I added seven comments. Can we please address these comments?”

[19:38] Brad: And it’s really nice to have one app where you ask a question, comment on code, build cool things. At the end of the day, it’s one landing point to give you all that in one apparatus.

[19:50] Brad: And one could kind of say, “Oh, it’s too much.” But as I’ve been using it, it feels like the right amount of tools and equipment in one packaged app.

[20:02] Brad: Once I started doing this, I haven’t gone back, which is kind of why I’ve switched from the CLI to the app and haven’t looked back.

[20:05] Brad: There is no automatic worktree support. There is no amazing review built into the CLI. And to be fair, the CLI is still excellent, and you can still use it. It has different use cases.

[20:17] Brad: But for me personally, using plugins, using this review pane, and I guess the one we talked about previously was automations—where you could run things scheduled on your computer using a basic prompt—all of that stuff is unlocked through the app, and I think it only continues to grow.

[20:34] Brad: So if you have not given the app a try, please give it a try and let me know what you think.

[20:37] Brad: I think the usage limits for Codex are also much more generous than Claude Code. You won’t hit limits because the models are efficient and capable.

[20:47] Brad: And I think, honestly, what I would do—if you have a great idea, like you want to build an app, you want to build a website, you want to do something—download Codex, sign in, just chuck two or three paragraphs of an idea in there and just let it go.

[20:59] Brad: Choose GPT-5.5 on extra high reasoning. So on the bottom right of the chat interface, you can choose your model and you can choose your reasoning level. Choose 5.5 on extra high and just let it go.

[21:10] Brad: Oftentimes it’ll work for 30 minutes, 45 minutes, where back then, in 2025, Claude Code would work for one minute and that would be amazing. I thought, “Holy crap, one minute of an AI doing what I told it to do and doing it pretty well—that’s incredible.”

[21:28] Brad: Now I go much bigger. I do parallel. I lean on plugins. I lean on these tools, and I can just get a whole lot done.

[21:34] Brad: I would like to caveat it with: it’s not perfect. So you’ll still have to do a little bit of trimming, a little bit of cleanup. Who knows, maybe in two years it’ll all be solved. But right now, you’re more powerful, more productive, and more connected.

[21:47] Brad: It’s easier, I think, too, to approach these things, given that it’s inside an app. So all things said and done, we’re in a future where your idea is almost the limiting factor.

[21:58] Brad: If you have a good idea, Codex can probably do it, or get you really damn close to where you’d be at. So if you’re an idea person and you have stuff cooking, I would definitely give it a try. You will not be disappointed, I imagine.

[22:11] Ben: Yeah, no, it’s funny on the plan mode thing. As you were talking about that, I was kind of thinking too: I haven’t used plan mode in forever. And I fully am with you that it’s just not necessary anymore.

[22:24] Ben: And it seemed like back in the day, if you got the initial action of the AI wrong, it was hard to kind of bring it back in and course-correct. And you were better off just starting a new chat and trying to be more specific.

[22:35] Ben: And I think there have been lots of studies and people talking about how one-shotting is actually better, oftentimes, at least with the older models. One-shotting was actually better than trying to continually go back and forth with it because its performance would deteriorate the longer and longer the message history went.

[22:53] Ben: I’m sure now that’s probably less of an issue. But yeah, usually it’s just like you don’t even need to do the plan mode. You’re just kind of like, “This is what I want.” And they always do a pretty good job of asking follow-up questions too if something isn’t very clear.

[23:05] Ben: It’s like, “Well, what do you think about this?” or “How do you…” It’ll give you recommended paths.

[23:11] Ben: And then, yeah, as you were talking, I was doing the review thing, which was pretty cool, which I hadn’t seen before, where you can go in and make the actual local edits and stuff like that.

[23:21] Ben: And then also too, there are so many plugins. I was looking and I was like, geez, there are so many. There are things that I use pretty often where—it’s funny now because it sounds so old school—but I still would consider myself pretty technically advanced from an accounting perspective.

[23:39] Ben: And there are things that I have Python scripts for where now I’m just like, “Why don’t I just connect this plugin?”

[23:46] Brad: Yeah, and even on that point, if you have a plugin but you had a Python script to do it, my guess is Codex would use that plugin and create its own Python script to get things done.

[23:59] Brad: So in the era of high intelligence, oftentimes Codex just reaches to write a local script that you never even see, but it gets the results that you want.

[24:08] Brad: And on the long chat threads, that was also a big issue with Opus and even Claude Code, which I totally forgot to bring up, is that now we’re in an age of long context.

[24:20] Brad: As you run out of context, there’s this compaction, which kind of summarizes what you did before and pushes you forward. That stuff works so well now. I’m blown away at how well it’s able to retain a long chat thread.

[24:33] Brad: So I think, to add on one last note: no plan mode, long chats, ambitious projects, use these excellent plugins, automations, or review tools, and ensure you’re using 5.5 on extra high.

[24:46] Brad: The long chat threads used to be kind of a skill. You used to have to know when Claude Code was sucking and start a new thread or clear things. And it had auto-compaction, but it wasn’t as good.

[24:56] Brad: Now we’re in the age where, like I said, truly the idea is limiting because these tools have gotten so good at understanding your intent, carrying your long conversation, enabling data, that you don’t have to worry about half of these things anymore.

[25:10] Brad: Which actually took skill and knowledge, because when I was on Twitter all the time last year with all the AI tooling, I would learn tips and tricks on how to use Claude Code better and better and better.

[25:20] Brad: Now it feels like the barrier to entry in using Codex effectively is really low, which is good because you shouldn’t have to be on Twitter to understand how to wield these tools. And it shouldn’t be that hard to just get things done.

[25:31] Brad: So yeah, thanks for bringing that up. I completely forgot. I think a lot of these things have changed so rapidly, and AI evolved so fast, that you just expect it to work and you don’t think about what it was two months ago even.

[25:41] Brad: I think long threads is one that, in particular, has felt very solved to me, at least in the Codex use case.

[25:48] Ben: Yeah, it’s fascinating. So I talked about it on this podcast, but I have my own accounting AI use-case YouTube channel, Augmented Accounting.

[25:58] Ben: And I was just thinking—I remember when my first video on that channel was. It was like seven months ago from now. And in that video, I was using LangChain because when you were building an agent at that time, there were a lot of things that you kind of had to manage a bit more hands-on.

[26:17] Ben: Like a long-running chat history, or just—yeah—one of the main things I remember being difficult was trying to manage how long your conversations should be, like how much the agent should have in its memory, because otherwise it starts degrading if you have too much.

[26:35] Ben: And it’s just funny looking back on that video seven months ago to now—you don’t have to think about that. It’s absolutely insane how fast it moves.

[26:47] Ben: And yeah, on the plugin stuff too, I mean, there’s one that I was looking at where I was like, “Wow, why am I doing what I’m doing?” There’s Pipedrive, which is a CRM.

[26:57] Ben: And I was trying to figure out the other day, “Okay, Pipedrive API,” and I needed to think about making a script to kind of get my data out. And I’m just like, I need to just use the plugin and save myself all the brain damage from that.

[27:09] Ben: Now, I was using Codex CLI to kind of think about building the Pipedrive API connection, obviously, but—or building the connection with the Pipedrive API.

[27:17] Ben: Or SendGrid. I’m doing this live. This is not scripted. The SendGrid plugin—I mean, that email crap was always so painful, and the fact that there’s just a plugin that can kind of help you deal with emails when you’re building an app is very nice.

[27:34] Ben: So yeah, that’s crazy.

[27:40] Ben: And I guess let me ask you a question live, and it’s okay if you don’t know. When you make a plugin, can you make it where permissions are locked down? If I wanted to create a plugin and share it with only you, Brad, is that possible? Or is it open to everybody if you make one?

[27:58] Brad: That is a good question. I don’t have the answer, but I would hope that there is a way to do that.

[28:02] Brad: There is a plugin marketplace, which is kind of shown by default when you open up the plugin surface. These look to be OpenAI ones and very large brands. I’m sure there are plenty of other plugins that are not here. I’m not sure how to get to them or if it is a managed marketplace. I don’t have an answer on that.

[28:23] Brad: At some point, I like having it be a curated list, but at the same point, it’s fun to share things.

[28:29] Brad: You have MCPs. Another thing we didn’t even talk about was skills, which we have talked about before. But skills, plugins, MCPs, automations—all this stuff feels like it’s working together well finally.

[28:41] Brad: I think oftentimes we had bits and pieces here and you had to fight the tooling, but skills are just another layer on top to say, “Hey, go use this plugin, go use this MCP,” and kind of embed all these instructions in a need-to-know manner.

[28:58] Brad: Now it’s all here front and center. I think there is a lot of effort to get here across all the AI coding tools—lots of different standards, et cetera, et cetera. Now we’re getting closer. It’s a much better system.

[29:09] Ben: Here’s what’s super interesting too. I’m looking at the skills. So I’m in Plugins, Manage, and I’m looking at my skills. And it picked up some skills that I had in Claude that I never created in Codex.

[29:19] Ben: It was just like—it must have been when it installed—it just saw skills that were available in other folders. I’m not sure how that works, but I think that’s what they do.

[29:27] Brad: Yeah, I think they’re trying to reference it all because, yeah, at this point it’d be nice if some of the AI tools cooperated a bit better, and I think that’s part of it.

[29:35] Brad: So yeah, you have plugins, you have apps, you have MCPs, you have skills. Sometimes it gets a bit confusing, but at the end of the day, a lot of this time it’s just packaging up either valuable data or valuable instructions.

[29:47] Brad: And the AI agent is really good now at knowing when to use these and how to kind of coordinate and wield them.

[29:53] Brad: Previously, MCPs used to blow your context window and you’d be shot. Now it just feels like I don’t even think about it. And that part is a true unlock.

[30:02] Ben: Yeah, I agree. Cool. Well, should we wrap it up there, Brad, and hit it with our bookmarks?

[30:07] Brad: Yeah, let’s do it. I need to pull up my bookmarks.

[30:14] Ben: Oh man, I know. We were in the middle of that convo, and I was like, “Oh shoot, we’ve got to do our bookmarks.” I do have one, actually. Oh God.

[30:19] Ben: Okay, okay. So mine—and this is not planned at all—but I have bookmarked something from the Startup Ideas Podcast, shout-out to Greg Eisenberg.

[30:33] Ben: He did an hour-long video with someone named Riley Brown called “How to Use Codex: The Codex Masterclass.”

[30:40] Ben: So again, this was not planned for this episode, but I do have it listed because some of the things that Riley supposedly talks about—just seeing the sizzle reel and what they were saying—was creating videos with Remotion, which is really cool.

[30:50] Ben: And then I think the other one was using the image tools because obviously OpenAI released a new image model that’s supposed to be really fantastic, and that’s something I’m kind of interested in.

[31:01] Ben: Like I just talked about using Canva, but should I just be using the image model to make—if it’s like a brochure and a flyer—should I just be using the image model, or should I still be using Canva’s MCP?

[31:15] Ben: So I have that bookmarked to go give it a watch. I started it, and I say this to commiserate with Greg because we have a podcast too, you and I, Brad, and we do YouTube videos.

[31:29] Ben: And the YouTube video—I think the editing messed up—where half of the Codex sidebar was blocked by their heads. And so people in the comments are complaining like, “Oh, we can’t see the sidebar because you guys’ heads are blocking it.”

[31:44] Ben: So anyways, long story short, I feel bad for him on that. But yeah, I plan to watch the whole thing. Started a little bit of it, but I’m even more inclined to do it after our talk because yeah, this is pretty cool stuff.

[31:57] Brad: Yeah, we’re on the frontier. Cool.

[31:59] Brad: Mine is from Evan Bacon. So he’s the head of AI at Expo. They’re kind of the React Native meta framework. But Evan shared a new tool that actually integrates with Codex.

[32:09] Brad: So his is called serve-sim, which essentially can show the iOS simulator in the right pane of Codex.

[32:18] Brad: So that right pane of Codex can show you a summary, it can show you a review, and the third thing it can do is a browser.

[32:24] Brad: And so what Evan had hooked up is basically, you navigate your browser to localhost port 3200. That’ll then actually show your simulator in the browser in Codex so that if you’re working on iPhone apps, you basically have the code and the app side by side, which oftentimes is not very easy.

[32:42] Brad: And this is very fresh. So as of recording, this came out yesterday and a bunch of people are posting about it. I can’t wait to give it a try.

[32:49] Brad: It feels like now that we have this app surface, people are spending more time building cool tools to fit into the Codex app. And this is a pretty cool addition by Evan. And he works on a bunch of cool stuff there.

[32:58] Brad: So I’ll share it in the show notes, and honestly, we’ll probably give a little review in the next podcast once I give it a try. Once I saw it, I thought, “Damn, that’s a good idea. Why don’t we have that?” And those are the best ideas.

[33:11] Ben: Nice. Nice. Very cool. Awesome. Well, good stuff, Brad. I think we covered a lot of good stuff.

[33:16] Ben: And yeah, I’m excited. I’m going to go plugin-crazy here after we drop and fully automate my life with some of this stuff.

[33:22] Brad: There is a lot there. Take your time. But hey, that’s part of the learning process.

[33:27] Ben: I do feel like—I’ll leave us with this—I feel like, you remember that movie Click with Adam Sandler? Do you remember that?

[33:37] Brad: A little bit, I think.

[33:37] Ben: Okay, well he has a remote and he can fast-forward parts of his life. But anyway, not to do the whole plot. I feel like with AI stuff, it’s like, “I’m going to have AI do this. I’m going to have AI do that.” And all of a sudden, you’re just not doing anything anymore—you’re just managing. And that’s how I feel.

[33:52] Brad: That is our future. I love it.

[33:56] Ben: Cool. Awesome, Brad. Good stuff. We’ll do this all again next time.

[34:01] Brad: Awesome. Sounds good. See ya.

[34:03] Speaker: Thank you for listening to the Breakeven Brothers podcast. If you enjoyed the episode, please leave us a five-star review on Spotify, Apple Podcasts, or wherever else you may be listening from.

[34:13] Speaker: Also, be sure to subscribe to our show and YouTube channel so you never miss an episode. Thanks, and take care.

[34:17] Speaker: All views and opinions by Bradley and Bennett are solely their own and unaffiliated with any external parties.