How Dan Shipper Runs His Whole Business on Codex and GPT 5.6

Dan Shipper does his email without opening his inbox.

His company, Every, runs multiple products (Cora, Proof, Monologue), a consulting arm, and a media business. And he manages most of it from one app: Codex, running GPT 5.6.

Not ChatGPT. One agent with access to his whole computer.

Greg Isenberg had him on to screen share the entire setup, and it's one of the most useful walkthroughs I've seen this year. Dan doesn't talk theory. He shows his actual inbox, his actual company feed, and then builds a small SaaS live on camera.

I took notes so you don't have to. Here's the full breakdown.

Codex is not a coding tool (that's the whole point)

The biggest mistake founders make with Codex is in the name.

"People think it's Codex, so it's for coding," Dan says. "And they're like, what, I already use ChatGPT. It's totally different. It's not the same thing at all."

His actual words: "Codex with 5.6 is my operating system for work. I do everything from my emails to writing to training models."

On where GPT 5.6 sits in the model landscape, Dan is blunt. Claude's Fable is the S-tier model, "so good it's illegal," but it's slow, expensive, and you need real skill to use it well. GPT 5.6 is the A-tier daily driver. His analogy: Fable is a tactical nuke, 5.6 is a Porsche. Great to drive around town, handles well, works for everything.

His estimate: 5.6 gets you 60 to 70% of Fable's raw power, at a fraction of the cost and much faster.

For most founder work (email, research, planning, standard code), that trade is a no-brainer.

His inbox is a feed of cards, not a list of emails

The first thing Dan shows is Tend, a small app he built for himself inside Codex.

Every day it sweeps his inbox, finds every unread email, and turns each one into a card: a summary of what's inside plus a draft reply, already written in his voice.

He doesn't write emails anymore. He reviews them.

"For 90% of emails, which are scheduling stuff and more generic, it's really good," he says. "It's as if I wrote it."

Greg's reaction was the best summary: it's how wealthy executives answered letters in the 1960s, Mad Men style. A secretary reads everything, drafts the reply, you approve. Except now everyone can afford the secretary.

Two details make this work:

→ The system compounds. Every archive, edit, and reply feeds back into a "feed policy" prompt: what Dan cares about, what to surface, what to skip. It gets better every day.

→ 5.6 made it viable. "5.5 was fine but not quite good enough. 5.6 makes far fewer mistakes. It doesn't send emails I wouldn't want it to send."

If you run any kind of pipeline (sales, hiring, support), this pattern transfers directly. It's the same logic behind the lead nurturing strategies that good sales teams automate: stop handling every touch by hand, build the system that handles the touch, then review the output.

Pulses, router threads, and an email address for your agent

Email is just one feed. Dan has a "pulse" for every part of his life he cares about.

His company pulse pulls every Slack message and meeting note he hasn't seen and turns them into cards. He scans the feed, decides what's worth pinging someone about, archives the rest.

Then it gets weirder (in a good way).

He built Mailroom, which gives his Codex its own email address using the plus trick: dan+codex plus a random string. A router thread checks that inbox every 5 to 15 minutes and routes whatever comes in to the right pulse.

So when someone at his company needs an ops task done, they don't email Dan. They ask the internal Slack agent, and the Slack agent emails Dan's Codex. The work happens, Dan just sees it happening.

There's also Chronicle, an opt-in feature that takes local screenshots of his screen and turns them into a context feed. Creepy on paper, useful in practice: the agent knows what he's working on without being told.

And a "record and replay" plugin: you say "watch me do this," do the task once, and it turns your recording into a reusable skill.

The pattern under all of it: every recurring input in your business can become a feed of cards with a suggested next action.

They built a SaaS live, and the idea behind it matters more than the app

Mid-episode, Greg and Dan decide to just build something.

The insight they start from: anyone can one-shot an app now. So the thing customers actually pay for is maintenance. Proof that someone is behind the product, fixing bugs, shipping.

So they build Turnaround, a badge you put on your site showing live maintenance activity: commits in the last 30 days, time from bug report to fix. A status page for the vibe code era.

Dan types one long prompt, hits go, and lets it run while they talk. The build uses two commands worth knowing:

→ LFG, from a plugin by Kieran Classon, who runs Cora: it loops through plan, execute, review, compound the learnings, and repeats until the thing is built.

→ Goal mode: give the model a long-term goal plus verifiable success criteria, and it keeps going until it gets there. Dan had one run going 4 hours during the episode. Over the weekend before, he ran two that went about 20 hours each.

Twenty minutes later they have a working mock: badge, landing page, GitHub ingestion plan. Dan's honest assessment: "We're 70% there." Design still needed work (5.6 kept insisting on warm paper backgrounds, which made me laugh), and he said he'd take the design to Claude and port it back.

Not magic. But 70% of a real product, built as a side activity during a podcast.

The Codex-native SaaS opportunity

This is the part I'd rewatch if you're building software.

Dan thinks the "SaaS apocalypse" narrative is dumb, and his reasoning is sharp. Yes, anyone can one-shot an inbox app. He literally did. But 99% of the world wants software maintained by someone else. His own team tried vibe coding their CRM, it became a mess, and they moved to a vendor with a CLI.

The opportunity he sees: SaaS designed to be used inside Codex, where the customer brings their own agent, with all their context, to your software.

The economics are the kicker. SaaS companies bolting agents into their products pay token costs on every seat. A Codex-native app doesn't run the agent, the customer's agent comes to it. "Your margins are back to being a regular software business instead of an AI business."

Greg pushed on distribution, and Dan's prediction: within a year, OpenAI and Anthropic will run app stores for agent-native apps. Greg remembers getting 25,000 downloads a day from an App Store feature in the mobile era. Same playbook, new platform.

Dan's advice if you're early: pick a category, build the Codex-native version, and start ranking for it now, so when it goes mainstream you're the default.

I covered Greg's earlier experiment with Codex running sites autonomously, and this episode is the logical next step: the tooling moved from "agent does tasks" to "agent is the workplace."

Pirates and architects (the new technical co-founder)

My favorite framework from the episode has nothing to do with models.

Dan splits AI-era builders into two types:

→ Pirates: explore fast, get things to 70%, figure out if the idea is valuable at all.

→ Architects: take a proven 70% thing and turn it into a polished, well-oiled machine.

Dan knows he's a pirate, so he deliberately pairs himself with an architect on every project.

His bigger point: "I need a technical co-founder" is an outdated frame, because everyone can code now. The real question is which temperament you are, and who covers your other half.

If you've ever shipped five prototypes and finished none of them, you already know which one you are.

Steal the setup, but start with one thing

Both of them land on the same warning, and it's the right one.

Don't watch this and try to rebuild Dan's whole system. "You'll just end up spending a bunch of time making a setup that doesn't work," he says. Start with one problem you actually have, solve it, build from there. Dan's system grew over years, one piece at a time.

And don't do it from FOMO. "It's much better to get into a new AI tool because you're genuinely curious. Not because you're afraid of being left behind."

The frontier move, if you want one: Dan says fine-tuning your own model is the next step after skills. His team is training a copy-editing model on hundreds of thousands of internal edits from six years of publishing, because frontier models still can't copy edit like their editors do. If your company has years of proprietary decisions in some niche, that's a trainable asset most founders are sitting on without knowing it.

That matches what I keep hearing from guests: the leverage lives in your data and the system you wrap around it, not in the model. Nick made the same point on the AI co-founder episode: the founders winning with agents are the ones who productize their own context.

FAQ

Is Codex only useful if you can code?

No, and that's Dan's core argument. He uses it for email, daily planning, company monitoring, research, and writing. The coding ability matters because the agent can build its own tools while doing your knowledge work, but you interact with it in plain English. His onboarding advice: download the desktop app (skip the CLI), give it access to your computer, and ask it "what could you do for me based on how I use this machine?"

How does GPT 5.6 actually compare to Claude's Fable?

Dan's live estimate: 5.6 is 60 to 70% of Fable's raw capability, but faster, much cheaper, and easier to use. Fable emptied his entire bug backlog on its own, "as long as you're willing to spend the money," while 5.6 still needs your thinking on hard bugs. For design taste, he still rates Claude ahead. His setup: 5.6 as the daily driver, the S-tier model for the hardest problems.

What is a "Codex-native" SaaS app?

Software designed to be used inside an agent environment, where customers bring their own agent (with all their personal context) to your product, instead of you embedding an agent in it. You skip the token costs, keep classic SaaS margins, and sell maintenance and reliability, which is what people pay for now that anyone can one-shot a v1.

Do I need Dan's whole setup to get value from this?

No. Both Dan and Greg say the opposite: pick the one recurring task that annoys you most (email triage, meeting notes, a weekly report), build a single feed or skill for it, and let the system grow from there. Rebuilding someone else's full workflow from a YouTube video is how you end up with a complex setup that doesn't fit how you work.

Hear these playbooks before they're obvious

Dan's edge is simple: he saw the shift a few months before everyone else and built for it. By the time "agent-native SaaS" is a LinkedIn buzzword, he'll already own the category.

That's the whole reason I run the Profitable Founder Podcast. Every week I sit down with bootstrapped founders doing $100K to $10M a year and get them to share their screen the way Dan did here, with the real numbers and the mistakes left in.