Greg Isenberg has never downloaded Codex.
Not once. He says it straight up in the first two minutes: "I am not on Codex today. In fact, I have never downloaded Codex."
The guy lives in Claude Code. Most of his content is about it. And yet his smartest friends kept repeating the same line: OpenAI's Codex is the super app for AI.
So he did something I respect. He called Riley Brown, the person most likely to change his mind, and gave him 64 minutes to do it on camera.
Riley's opening argument was one number: he works with 7 engineers, and all 7 switched to Codex.
I watched the full episode so you don't have to. Here's the whole Codex vs Claude Code debate, the parts that convinced me, and the parts that didn't.
The setup: a Claude loyalist vs a converted team
What makes this episode work is that neither guy is selling you a course.
Greg walks in skeptical. His exact framing: "By the end of this episode, I want to be converted. I might be like, you know what, I'm just going to stick with my own stack."
Riley walks in without hating on anyone's tools. He even says switching from hot tool to hot tool is the wrong strategy: pick a stack and stick with it. If Claude Code inside Cursor works for you, keep going.
But he's permanently moving to Codex anyway. His reason is boring and that's what makes it credible: his whole engineering team tried it, agreed it was better for how they work, and switched together.
→ That's the signal worth watching, before any feature demo: a whole team tried it, compared notes, and stayed.
What Codex actually is (and why one interface matters)
Riley's definition is the cleanest I've heard: Codex is a state-of-the-art AI agent on your computer, running the GPT-5.5 model in one interface that can build an app, create any document, control your computer, and run automations.
If you have a ChatGPT subscription, you already have access to it.
His demo list from the episode:
- A train simulator with working physics and a crash counter, built in one prompt (he posted it on Twitter the day before filming)
- Spreadsheets with charts, Word docs you can edit and export, PowerPoint decks that open directly in Canva
- Landing pages and lead magnets in the same chat where he does his knowledge work
His core claim: Codex is the only interface that does vibe coding and knowledge work in one platform. Cursor can build you an HTML file, but it can't just hand you a doc like a Claude artifact. Claude's desktop app can, but you have to go through the co-work feature, and Riley is not a fan of co-work.
As a founder, that pitch lands. I'm constantly bouncing between a code editor, a doc, a deck, and a spreadsheet. One interface for all four is not a small thing.
If you want the deeper version of this argument, I broke down the same thesis in my piece on the Codex operating system. Riley's episode is basically that idea with screen recordings.
The Remotion play: launch videos with a million views
This was the segment where Greg leaned in, and honestly, me too.
Riley makes his product launch videos with Remotion, a tool that generates motion graphics as code. Anthropic used the same approach for their early product launches. Inside Codex, Remotion is a plugin: you type @remotion in any chat and it opens right there, CapCut-style timeline and all.
The numbers he dropped:
- Some of his team's launch videos have passed 1 million views
- One Remotion video he showed on screen was sitting at almost 800,000 views
- Greg knows multiple people who went from zero to 100,000 Instagram followers on the back of high-quality Remotion videos
Then Riley showed the part that makes it a system instead of a party trick. He built a custom skill called Internet Image Puller. It runs for about 10 minutes, pulls every asset a brand has on the internet (logo, colors, fonts), and drops them into an HTML file. Remotion then uses your actual brand kit in the video without you manually feeding it anything.
→ Launch video, on brand, from a prompt. That's a distribution weapon for a solo founder.
Skills, plugins, and the Friday 9am critique
Riley admits the vocabulary is a mess. MCPs, plugins, skills, integrations, connectors. His fix: clump them all together and call them skills. An agent without skills is just a generic model. Your job is to hand it the tools you already use.
His stack inside Codex: Slack (his company has around 30 channels, and he asks the agent to summarize everything that pertains to him), email, Notion, Sheets, and Expo for previewing mobile apps inside the chat.
One difference worth knowing: plugins need OpenAI's approval, but skills you can create yourself by literally opening a chat and typing "I want to create a new skill that does X."
The automation demo was my favorite part of the episode. Greg had Codex analyze his podcast and generate a report of what he's doing wrong. Then, in one sentence, he turned it into a recurring job: "Create an automation, 9:00 a.m. on Fridays." Done. Active. Next run scheduled.
Riley's rule for this: never build the automation first. Do the task once with the agent, check the output, then tell it "do this every Friday." Most people design automations backwards. This is the same principle that shows up in every good writeup on business automations: automate a process you've already validated by hand, not one you hope works.
The pattern Riley sees most in the wild? Founders having AI read their Slack and email twice a day and send back one summary with links. Boring, and it compounds.
(Side note: the report told Greg his interviews are way too soft and his guests get "a friendly runway." He read it on camera and immediately started grilling Riley harder. Worth the watch just for that.)
The honest catches: price, effort levels, and design
Greg took the AI's feedback and asked the hard questions. Riley didn't dodge.
Cost. GPT-5.5 is twice as expensive as GPT-5.4 through the API. Riley's defense: it finishes jobs with fewer tokens, so the per-task cost is closer than it looks. And on the $100 or $200 per month plans you get so many tokens that for most founders the API price is irrelevant anyway.
Effort levels. The people Riley trusts run 5.5 on low or medium effort for most tasks, because extra high will over-engineer a simple change and wander off doing unintended things. Riley admits he's not disciplined enough and just runs extra high over and over. There's also a fast mode, which burns credits quicker.
Design. This is the one that surprised me. Riley concedes the point that Claude is way better at design. His workaround: build the whole app in Codex, then tell Claude to design it. He even ran Claude Code inside Codex live on the call and had it restyle an app into a chessboard.
So the real verdict is a stack, plural. Riley runs Codex as the operating layer with Claude Code inside it for taste. If you're still weighing editors, my Claude Code vs Cursor breakdown covers the other half of this decision.
Browser agents are about 3 months away
The last third of the episode is about where this goes.
Riley has been skeptical of AI controlling browsers. Slow and a little boring, in his words. This episode is the first time he said "oh, it's actually starting to be faster." His prediction on camera: browser agents will operate as fast as humans, and it's about 3 months away, not a year.
He also made a founder-brain observation I keep thinking about. There are rumors of Meta and Microsoft asking employees to record their screens while they work. If super apps can learn workflows from recorded video, the value of recorded content goes way up. Greg's response: "just from a startup idea perspective, there's a lot there."
→ If your moat is a repeatable workflow, documenting it now is cheap. Training an agent on it later might be the payoff.
Steal Riley's starter projects
Riley's advice for your first weekend with Codex is a list of five concrete reps:
- Make a simple game, then have browser use play it against itself. You'll feel where agents are headed.
- Run a max-effort research task: one topic, organized into a spreadsheet, then turned into a doc and a presentation.
- Build a 3D simulation, purely for fun. (This is the train simulator energy.)
- One-shot a Swift mobile app from your existing website's code. You still need a Mac with Xcode, but Riley says the results got "actually scary."
- List everything you do daily, pick the most annoying task, do it once with the agent, then turn it into an automation.
Riley's bar for the whole exercise: one useful automation built in a couple of hours. Clear that and the tool has paid for itself.
FAQ
Is Codex better than Claude Code?
Depends what you're optimizing for. Riley Brown's team of 7 engineers switched to Codex because it combines coding, documents, spreadsheets, and automations in one interface. But Riley himself says Claude is way better at design, and he runs Claude Code inside Codex for exactly that. For most founders the usable answer is to layer them, the way Riley does.
Do I need a separate subscription to use Codex?
No. If you have a ChatGPT subscription, you have access to Codex. Riley points out the $100 and $200 per month plans include enough tokens that most solo founders never think about API pricing.
What is Remotion and why do founders use it?
Remotion generates motion-graphic videos as code. Inside Codex it works as a plugin you call with @remotion. Riley's team uses it for product launch videos, some of which passed 1 million views, and Greg knows creators who grew from zero to 100,000 Instagram followers with it.
What's the difference between skills and plugins in Codex?
Plugins are official integrations (Slack, Notion, Canva, Expo) that OpenAI has to approve. Skills are custom abilities you create yourself by describing them in a chat. Riley's Internet Image Puller skill, which gathers a brand's logo, colors, and fonts from the web in about 10 minutes, is a skill he built with a prompt.
I spend a lot of time pulling apart episodes like this one, but the better version is hearing bootstrapped founders explain their stack decisions themselves. That's the whole format of the Profitable Founder Podcast: real founders between $100K and $10M a year, walking through exactly what they built and why.
