In the spring of 2020, most of the world was watching Tiger King and feeding sourdough starters. Brendan Falk was in a San Francisco hacker house on his fifth pivot in six months, trying to find an idea Y Combinator would actually like.
YC had already accepted him and his co-founder, Matt. Then, on the same call, told them they didn't like the idea they'd applied with. Come up with something new.
That pattern repeats through Brendan's career: get in the door, find out the plan was wrong, keep building anyway. It took him from a command-line startup with hundreds of thousands of weekly users, to an acquisition by AWS, to a services company that didn't work, to Hercules, a seven-person company that he says is now doing tens of millions in revenue by letting people who have never written code build the software their businesses run on.
Codeship exists for builders who want to become founders. Brendan is a useful case study because he's candid about the parts that went wrong, including a sale he says he wouldn't make again. Codeship founder and CloudBees CEO Moritz Plassnig sat down with him at the Hercules office in San Francisco to talk about all of it.
This conversation has been edited for length and clarity. You can watch or listen to it in its entirety here or on YouTube.
The early days
Moritz: We’re here in San Francisco today and it’s hard to imagine you anywhere else, but you didn't grow up in SF. Tell us about your background.
Brendan: I grew up in Australia and moved to the US for undergrad at Harvard, where I studied economics and computer science. I fell in love with tech really young and was always building apps. I started a bunch of mini companies in high school and college, and straight after graduating I started my first real company, Fig, a developer tools company. We were acquired in 2023. Now we're running Hercules.
Did you always know you wanted to start a company?
I don't understand people who don't want to start companies. At every internship I had, I'd find the best person and say, "You're really good at this. Why don't you leave and start a company?" I couldn't comprehend why someone wouldn't. I've learned that's not how most people think. But to me, it's the most fun thing to do, it's how you can have the most impact, and it's the most interesting. As stressful as it is (it's like eating glass, as Elon Musk says), I still love it.
When I was growing up in Austria, starting a company just wasn't a thing. I didn't know it was possible. Were you encouraged as a kid, or did you know people who'd done it?
Both my parents are doctors, my brother works in finance, and I don't think anyone in my extended family is an entrepreneur. When I was a kid, I thought a company was like, Angry Birds. So I'd build an iPhone app. I wasn't thinking about boards and governance. I just liked building things. My family, my brother especially, describes me as being so annoying: "I have an app idea for this. I have an app idea for that." That was what excited me, and it turns out you can build a company around it.
"I'd be so annoying, like, 'I have an app idea for this. I have an app idea for that.' That was the thing that was exciting to me, and then it just turns out that you can build a company around it."
Fig and the command line
How did you end up in developer tools, building products for other people who build?
My co-founder at Fig, Matt, was my college roommate for three of our four years. We met in our first week. In high school he'd built an HTML renderer for the iPad, and I had literally used it to learn how to code. So I thought, this guy's legit, we have to work on stuff together. He was always the one coding and I was the stereotypical business guy.
I was hell-bent on getting into Y Combinator. I applied on my own once with a crappy app, and Michael Seibel grilled me for 10 minutes. I didn't get in. When Matt came back from a study abroad, we applied together for the Winter '20 batch. We applied with one idea, pivoted before the interview, and then got the call: "You're in. But we don't like your idea. Come up with something new."
We did the early admit program, so we got in in October 2019 but started in summer 2020, after graduating. Then we were in pivot hell for six months. Every time we pivoted, we went back to the command line to start a new project: set up a Heroku server, set up CI/CD, set up GitHub, set up npm. Paul Graham talks about schlep blindness: doing painful things so often you stop recognizing they're painful. Matt and I finally looked at each other and said, "Why is this so hard?"
When we'd pitched our earlier ideas to friends, they were nice about it: "Sounds interesting, tell us when it's out." When we asked about the command line, they hated it. "I did an internship at Google and I had to use it." They don't teach it in school. It's expected knowledge. It had barely changed since the 1970s, and every engineer still had to use it.
"About 1% of people love the command line and think it's the best thing in the world. 10% are indifferent. The other 89% of people absolutely hate it."
So the question became: how do we make the command line easier for individuals, more collaborative for teams, and more secure for the enterprise? That was Fig.
So you were in YC during Covid then. What was that like?
We were the first remote batch. We moved to San Francisco in June 2020, and Matt and I lived together while we figured out what to build first. This was pre-AI, so building things just took time. Around October 2020 we made a big call: the first feature would be autocomplete for the command line. You had IntelliSense in your IDE and autocomplete in Google and Gmail, so the value prop was clear.
We liked it for two reasons. First, you either use it 100% of the time or you churn. Nobody's a passive subscriber. Second, you use it every day, so every time you open the command line, we can upsell other features. It felt like the biggest call ever, and it was the right one. We got to hundreds of thousands of users on autocomplete. And so many products started there: Cursor, GitHub Copilot, VS Code with IntelliSense.
Selling Fig to AWS
Fast-forward to 2023, you sold Fig to Amazon. What's the story?
Here's a challenge for you. In 2022, how many individual developer productivity tools had more than $100 million in revenue?
Probably none. JetBrains?
Right, as an IDE. Before that, Visual Studio. Then GitHub Copilot, which had an unfair advantage through its partnership with OpenAI. Then Cursor. Now it seems like everyone's hitting $100 million. But at the time, we'd picked a really small market.
Our goal was never just individual productivity. It was team collaboration and enterprise readiness. The mistake we made was being too focused on the command line as the core interface. We built a secrets manager similar to HashiCorp Vault, though obviously nowhere near as extensive. Why use us instead of Vault? Because we were the best for the terminal. But you don't just need secrets in the terminal. You need them in production apps and in CI/CD. That interface was already eroding, and we were too opinionated about it. And the people who love the command line don't want an app like ours. They want Vim and Neovim and some Arch Linux distro with all their customizations.
We had hundreds of thousands of weekly active users who loved the product, a real cult following. Then ChatGPT came out and we're like, “Holy crap, the entire world just changed.”
We'd been at it for three and a half years. We'd raised money and had to build a business. We were struggling to monetize individuals (we didn't want to charge everyone $50 a month), and the team product wasn't working because we were too focused on the command line. So we decided acquisition was best for the investors, the employees, and the company. We had an inbound offer we'd always brushed off, used it to run a process, and landed at AWS.
How we landed there shaped me as an entrepreneur: what outcome did AWS want? Copilot was eating everyone's lunch. Amazon had CodeWhisperer, which was technically strong, but AWS is known for infrastructure, not developer tools or beautiful interfaces. They needed three things: differentiation from Copilot, which the command line gave them; users, and we had hundreds of thousands of weekly active users; and a team that could build dev tools. We ticked all three boxes.
Within three months we'd re-architected the product to be AWS-ready and launched it the week before re:Invent. Today it's effectively Amazon's version of the Claude Code CLI, called Kiro CLI, and it's doing phenomenally well. But it was hard to see that coming at the time. It felt like Copilot would own everything and OpenAI would own everything else.
If you could go back knowing what you know now, would you still sell Fig?
Oh, no. Absolutely not. We had hundreds of thousands of weekly active developers, and now look at the Cursors, the Cognitions. Claude Code started on the command line. We would have moved beyond it and built IDEs, but knowing what I know now, absolutely not. At the time, I strongly believed we made the right decision.
The detour
Did you quit Amazon and start something new the next day?
No. As soon as I realized I wanted to start another company, I wanted to find my own idea, and I wasn't getting that on the team we'd been acquired into. I wanted a team where I'd be exposed to a ton of customers and problems, on the front lines, and also exposed to AI. So I did something that makes no sense on my LinkedIn. I joined the AWS Private Equity Business Development team.
I didn't know you did that.
Everyone has that reaction. The team works with 100 to 200 major PE firms and tens of thousands of their portfolio companies. I was the global gen AI lead. My job was to get parachuted into these firms and companies and help them come up with an AI strategy.
PE companies move really quickly. If I'm parachuted into a large enterprise, it takes six months to get anything done. And PE doesn't discriminate: big companies, small companies, every industry, every geography. If I was interested in insurance, I'd do a deep dive on AI for insurance, then move on to customer support. I worked on engagements across every industry, geography, and company size.
The big thing I saw, in 2024, was that the time to build software and agents was about to drop dramatically. That wasn't obvious then.
"I was talking to these CIOs, and they were going, 'I wish we could do X.' And I'm thinking, 'Give me a couple of engineers and a week and I can probably save you a million dollars.' It's an amazing moment in time, and they don't realize it."
I realized there was far more opportunity outside dev tools than in selling to other developers, because those people had no idea this was possible.
So when did you start Hercules?
I spent a year running our product at Amazon, then six months on the PE team. The company we started first was called Zeus, which is now Hercules. I'd worked with enormous enterprises. I'd seen how painful their governance issues are, how much money they have, how big the contracts are. The largest software companies sell to enterprises, so I knew I had to crack that. Rather than work my way up from SMB, I'd jump straight in.
The hypothesis was something like Palantir. We'd come in with our own stack, automate building software and AI agents for these companies, save them $10 million, and charge $1 million for the privilege.
What I got right was that the time to build software is collapsing. What I got really wrong is that the rest of the process isn't automated by AI. You can't cold email large enterprises; you need a warm intro. Discovery and requirements gathering need a human on the phone. Data cleaning is the same: It doesn't matter what's in the docs. It matters what's in Carol's head, who's been there for 10 years and has never written it down.
Ongoing maintenance is different too. Everyone said Accenture was dead. There's still so much human interaction these companies need. We were also very agnostic, trying to build everything, and the lack of focus was my mistake. It wasn't going to be a software business. It was a services business.
About six months in, we stepped back. We still believed the time to build software was collapsing. Lovable, Cursor, and Replit were blowing up. So we went back to what we did best at Fig, bottoms-up and self-serve, but instead of developers, we focused on enabling non-technical people to build software. That was August 2025.
Did you start Hercules with the same co-founder as Fig?
Grant was our first engineer at Fig, so I've worked with him for five or six years. He's great. Matt is now at Cognition, leading an AI product there. Pretty much everyone we used to work with is now at Cognition or with me. We're all still in the dev tools world.
Building Hercules
Did it worry you that other companies were building similar products at the time, or did you see that as proof it could work?
Plenty of investors were concerned. But we'd already raised from Andreessen, and they've been fantastic and supportive the whole way. Lovable and Replit had been working on this for over a year, so we were definitely behind. But when I looked at their apps, I saw gaps. Pricing was flat and chat-based, not usage-based, which wasn't going to scale. The models weren't there yet for long-running agents. We were trying to serve completely non-technical people. I think about my parents, who are incredibly smart doctors who've never coded, and even Lovable and Replit felt too complicated for them. And those platforms couldn't build production-grade, full-stack applications.
Everyone talks about AI app building as a market. It's not a market. It's a solution. What are the markets? What are the outcomes people want to build?
It's a means to an end.
Exactly. There are individuals building hobby apps and personal websites. There are prototypes. There's e-commerce on something like Shopify. There are landing pages: I'm a dental clinic, I need a landing page. There are internal tools. Those are all different markets.
We focus on large operational systems of record. Our customers aren't building landing pages or hobby apps. They're building CRMs, ERPs, HR systems, warehouse management systems. Vertical SaaS that's now hyper-custom to the business. Think of the most obscure industry you can. Does software exist for it? Probably not, because it's too small. Now a customer can build custom software for their business.
That's technically hard, and it has to be enterprise-grade. That's why I like the problem. We're not just the engineer. We're the engineering manager, the product manager, the accessibility reviewer, the legal and privacy compliance team, and the AppSec engineer. We're your entire outsourced software development agency, not just a coding assistant for developers. And because we serve large enterprises, it's not enough for the platform to be secure. The application has to be robust. You can't have your identity and access management system break in your application. That has to work.
All the “boring” stuff.
Exactly. But it's not boring if you need it to work. Large companies use us because our platform is secure, the apps we build are secure, and we're just as easy to use as everyone else.
Breaking through the noise
So you were selling to real, large companies even early on. Did you build a sales team, or is it still bottom-up?
It's all bottom-up. We still have seven people at the company, and we're doing tens of millions in revenue.
The hardest part is getting someone non-technical to believe we can actually build these applications. And with AI, that's expensive. I can spend two weeks on Canva for free and then decide to pay. We can't give everyone $100 of credits on a free tier. We have to get them to the aha moment quickly. And for the first 10 messages on any of these platforms, everything looks similar. Lovable builds a beautiful prototype with no functioning backend. Replit has a functioning backend, with its own constraints. So the challenge is showing customers we can do it, and then actually doing it, which takes a lot of context engineering, harness engineering, and digging into what each customer is trying to build.
There's so much noise right now. How do customers find you?
When I started, I planned to follow the Fig playbook: launch cool stuff on Twitter and everyone talks about you. But Twitter is crowded with every lab and company launching features every day. And our users aren't on Twitter.
Here's an example. We have a customer in Florida who runs a freight forwarding company. He coordinates planes and ships. He does about $20 million in revenue and has 350 employees, and he built his entire ERP on Hercules: supply chain, warehouse management, logistics, control tower, invoicing, and a mobile app. He'd never coded in his life. His revenue has gone up massively. And he's not on Twitter at all. He built his entire ERP on Hercules.
When Twitter wasn't working, I decided I didn't care about making money yet. I just wanted to build the best product, because if we could crack this, the ROI is 1,000x: going from paying a systems integrator to doing it yourself. So we ran Google Ads in really low-CAC regions and let Google work it out. It pattern-matched to India, Africa, and Southeast Asia. Those customers have the same businesses (construction, wholesalers, warehouses), with much lower willingness to pay. If we got a ton of feedback from them, we'd have the best product.
Vibe coding had just taken off, and there were hundreds of thousands of searches for "AI app builder." We could bid a cent for searches from those regions. We got a ton of users quickly. A lot of them churned, and we asked every one of them why. We gave away credits aggressively just to get feedback. That made the product much stronger and taught us who our customers were, what they were building, and why.
Then some of them started paying, and we thought, let's target developed economies too. The CAC payback was phenomenal. When that plateaued, we tried Facebook. Maybe 100,000 people a month search for "AI app builder," but a billion people a day use Facebook's apps. And instead of targeting a search term, we can target "you're building an ERP, you're in this industry."
Will that last forever? Absolutely not, and we're building other channels. But because we focused on the product, customers love it and talk about it on LinkedIn and Twitter, which drives SEO and AEO. If you ask ChatGPT, "I'm a business and I want to build software, what should I use?", it'll tell you Lovable and Replit do a bit of everything, Claude Code is too technical and expensive, and Hercules focuses specifically on this use case. That focus has been really good for us.
Wouldn't you worry that what you learn from those markets doesn't transfer to customers in the US or Western Europe?
The product lessons transferred. What we didn't build initially was the enterprise security: two-factor auth, GDPR, access controls. But we already knew that from Amazon. The go-to-market didn't transfer, so we still had to solve that separately.
Earning trust
Take the freight forwarder example. He found a website, probably never met any of you, and now runs a business he spent years building on your platform. Why do customers trust you?
A few reasons. First, the product experience is very different from something like Claude Code. Those tools are optimized to hit a benchmark: "Build a CRM," and it runs for five minutes and builds a CRM. We ask, "What are you trying to build? What company do you work at?" We ask a lot of disambiguating questions.
Then we break the work into milestones and build one at a time. If someone at OpenAI heard that, they'd say it sounds crazy, but it works for us. You could one-shot the whole thing, but you'd probably get it wrong. Every builder knows that even with all the planning up front, you want to be agile, not waterfall. Customers don't know what they want until they see it.
Developers went through a progression over two years: autocomplete, then a chatbot, then the chatbot writing to a file, then to multiple files, then running for 24 hours. If you've never used an agent that runs for 10 hours, that doesn't make sense to you. So we work customers up to it. For a CRM: phase one is a shell, phase two is the contacts page, phase three is opportunities. Every milestone is something they can use and react to. Because we are actually slowing them down, we're earning trust as we go.
Fifty prompts and 10 milestones in, they can say, "Just skip ahead and build the whole thing."
Then there's trusting us enough to migrate. Many of our customers spend hundreds of hours on Hercules in the first couple of weeks. We keep calling our shots and hitting them. We make mistakes, and when a customer emails for help, we don't say, "You're on your own." Three of our seven people focus on customer support and operations. In the early days the models weren't there. As they've improved, our users have gone up and issues per user have gone down, but we're happy to throw humans at the problem in the meantime.
We pitch ourselves a bit like a software development agency: you build it yourself, and we're here if you need us. If you're really abusing it, we'll charge for some hours, but we're happy to subsidize technical architecture questions. And customers are comparing us to an agency quote of $250,000 for six months of work that would probably fail anyway. So even before we've earned their trust, we're far cheaper. But ultimately, the product is really good. People can build what they want to build.
Having real people help when something goes wrong is powerful, especially when so many companies are automating support away.
I've told my team that customer support is the most important part of any company, because it's the front line with your customers. Every morning, every engineer, me, and the support team do a 30-minute standup on the tickets that came in overnight.
A real example: a customer wrote in, "You told me to use Resend. What the hell is Resend?" You and I know what Resend is. A freight forwarder doesn't. He doesn't know what an API key or a secret or a DNS record is. He just wants to send an email from his app. The first time, we walked them through it. The second time, we said, "Why don't we build Hercules Email, our own email service, embedded automatically in every app?"
"We help one customer, and then we solve it, so it helps the next 10,000 customers."
That's why we've built basically our entire stack and infrastructure ourselves. It's the only way to solve so many of these problems.
The SaaS apocalypse
It sounds like these customers were using vertical software, manual processes, or expensive contractors, and now they're building with you. Which brings me to the SaaS apocalypse. What's your take?
Very strong opinion. The narrative was: why pay Salesforce? Chamath started a company with the pitch, "We'll give you software that's 80% as good for 90% lower cost." So instead of paying Salesforce $1 million a year, pay for a cheaper version at $100,000.
We've seen the complete opposite. Our customers are happy to pay us more than they paid for their existing software, because it's not about saving costs. It's about generating revenue. If your core operations run much faster on software designed for your business, your revenue goes up. We've had customers turn off SAP, Salesforce, vertical software, and spreadsheets. We go to software categories that already exist and say, "Why not rebuild it for your business?"
I think Salesforce will still be really successful, and a new breed of startups will be too. We're not counter-positioning against anyone. Our pitch is this:
"For the last 30, 40 years, you have run your business working forwards from the way a software provider said you should run it. Now the cost of software is dropping so much. Doesn't it make sense that your software should work the way you want to run the business?"
You never built custom software because it was too expensive, too slow, and too risky, and you didn't know how because you're not a software business. We believe custom software is the revenue-maximizing strategy for every business. The question is still whether you can wave a magic wand and get there. But if models keep improving and, say, the cost goes to zero in five years (I don't think it will, but say it does), every business should run its core operations on custom software. We think of ourselves as a framework for customers to build mission-critical software, the same way developers plug into infrastructure tools to run custom workflows.
We'll probably cut into the systems integrator and implementation market more than SaaS. Software is around $500 billion. Systems integration, implementation, and services are another $500 billion. That's $1 trillion, and I think it all expands massively.
One last example. We got a support email:
"My pig inventory isn't working."
The whole team was like, what? It turns out it was a restaurant that sells pigs on a spit, in several sizes. They found Hercules through ChatGPT or something similar and built inventory management software specifically for pigs on a spit. That market didn't exist before. There was inventory software, but nothing hyper-custom to that business.
So my take: Salesforce can win, the systems integrators can win, and we can win. There's just going to be a lot more software, and it'll be much more custom to each business.
What builders should take from this
Broken down in hindsight, Brendan's career is a series of bets, most of them partly wrong, each one teaching him something the next one needed.
Fig taught him that a tool people use all day is a platform, and that loving an interface is a liability. AWS taught him what enterprises need and how slow they buy. Zeus taught him that cheap software doesn't make the rest of the work cheap. Hercules is what he built with all of it: a self-serve product for people who will never open a terminal, held to the standard of software a 350-person company can run on.
That's the kind of story Codeship is here to tell: the tradeoffs, the calls that looked right at the time, and what someone actually learned from building. Brendan was generous with all three.
Learn more about Hercules: https://hercules.app
