Matt Palmer on Replit's speedrun to $100M ARR
Matt Palmer on Replit's speedrun past $100M ARR since launching Agent, why the company's docs no longer speak to its own ICP, and why he thinks the best Replit users look a lot like the scrappy people who build incredible things in Google Sheets.
Matt Palmer works on developer education and marketing at Replit, which went from roughly $10M to $100M in ARR in under a year after launching its AI coding agent last September. We dig into what actually drove that hockey-stick growth, and the much harder problem underneath it: Replit's own documentation, brand, and messaging were all built for professional developers, and its fastest-growing user base increasingly isn't.
We get into why video has become central to how Replit teaches people to work with non-deterministic AI tools, the Diátaxis framework behind their documentation rewrite, and who Matt sees actually getting the most value out of the product.
The Q&A below is experimental — extracted from the episode transcript with AI. If you spot a mistake, let us know at jack@scalingdevtools.com.
Replit used to be squarely a developer tool. Now a lot of your users are semi-technical or non-technical. How has that changed the job?
Replit's been around since 2018 and was traditionally marketed developer-to-developer, as a place to learn to write code and deploy it fast — that's what I joined for. Since launching Agent, we see a lot of semi-technical or even non-technical users on the platform, which is both a challenge and an opportunity. I like the parallel to Neon, the database company: who needed a Postgres database a year ago? Developers building web apps. Who needs it now? AI agents building web apps for people and putting a database on the back end. A whole set of people are being exposed to traditionally developer-only concepts, like databases, for the first time, and it's my job to educate them alongside the more hardcore devs who are also getting into vibe coding.
Do those newer users even need to understand the underlying concepts, like what a database actually is?
That's a really valid question, and it gets at what developer education even means here. Nobody needs me to explain the difference between OLAP and OLTP databases. But they will care if I tell them: you want to build an app where people can vote on restaurants in real time, you'll need somewhere to persistently store that data, that's what a database is for, and agents are pretty good at writing basic data architecture and implementing it. You can stop at the outcome level — this is what you want, this is what you need to get there — and if someone's curious, they can dig into the nuances later.
You went from roughly $10M to $100M in ARR in under a year after launching Agent. What actually drove that?
I wish I had the full answer, but a lot of it is hard work, new technology, and the foundation the team built well before I joined — Amjad started the company in 2018 and positioned it to take advantage of leaps in the technology as they happened. About nine or ten months before this recording, Replit was a platform for building and deploying fast, but you still had to write the code yourself, even though all the infrastructure was there. When models like Sonnet started coming out, Amjad, who'd been thinking about AI-assisted coding well before that, made the call to lean hard into it, and we launched Agent in September. Since then it's just been a hockey stick. It wasn't obvious at the base of that curve — everyone was just putting in the work on something that was very much a hypothesis. The September launch itself was well received but still just a blip on the graph; the growth kept compounding faster than we could even celebrate each milestone, hitting $20M ARR and thinking that alone was insane before the product kept taking off.
When a product is growing this explosively, what actually matters in marketing, versus normal steady growth?
First, we made a huge pivot in what the product is and who we sell to, so our existing documentation didn't speak to our ICP anymore and our messaging had to change — and it's genuinely hard to shift people's existing perception. I still meet people at events who say, oh, Replit, I used that in college to learn to code. The most important thing from my side was getting our documentation squared away — the language we use, the frameworks we use to structure it. Then it's been adjusting marketing language and brand guidelines to cater to the right audience, while accepting that changing perception like that takes a long time. I asked Logan Kilpatrick a similar question a couple of years ago when he was at OpenAI, and he said basically the same thing — docs are what you end up racing to catch up on when a product takes off like this.
Tell me about Diátaxis and why you've put so much weight behind documentation specifically.
Diátaxis is a framework for thinking about how to structure documentation — separating reference material (how does this specific thing work) from how-to guides and tutorials (show me how to do this thing) — in a way that's useful, discoverable, and navigable for users. I actually feel grateful for the timing: three years ago, writing docs at this pace would have meant locking myself in a cave. Now AI can get us maybe 80 to 85% of the way there, which matters a lot on a fast-moving team that's shipping constantly. I'd also add a more personal point: very few people or companies are genuinely excited about writing docs. Doing that one thing really well, with real attention to detail, is a surprisingly easy way to stand out and prove you care about the developer experience, even though almost nobody treats it that way.
Why lean so heavily into video as a teaching medium, on top of the docs work?
Video is powerful, especially now that people are working with tools that are genuinely non-deterministic — you might run something twice and get a different result, or fail, adjust the context, and get something amazing an hour later. That's a very new phenomenon, for developers too, not just newcomers. Video lets you show someone I tried this and failed, but I'm going to try again, or exactly how I think through a problem, in a way text tutorials can't really convey — you can read documentation and still not understand someone's thought process for debugging a console error. It also scales in a way one-on-one help doesn't: I can put out one video that a hundred thousand people watch, but I can't get on a Zoom call with a hundred thousand people. And it's one of the most personal media available — lighting, pacing, tone, whether you crack a joke — all things a docs page just can't carry the same way.
Who's actually getting the most value out of Replit right now?
There's a broad spectrum — hardcore developers still get value, often for spinning something up quickly or scaffolding an idea they'll invest more time in later. But some of our real power users are semi-technical builders with a lot of agency — meaning the willingness to debug things, figure out why something isn't working, and push the boundaries of what's possible. I like the comparison to people who are really good in Google Sheets or Excel: scrappy people willing to sit in the trenches figuring out undocumented formulas, who've built genuinely amazing things in a spreadsheet. Drop those same people into Agent, and if they're willing to learn a bit and willing to fail and retry, we're democratizing the ability to create for them — not replacing developers, but recognizing there's a lot devs shouldn't have to spend their time on. The people getting the most out of it are the ones who've had ideas nagging at them for a while and can now just go build the thing.
What have you learned about what makes a good agent experience specifically, versus a good IDE experience?
My personal take is that a good experience for a developer tends to be a good experience for an agent too — you're constantly thinking about what context to give it and what to leave out to get the result you want, which is a lot like being a manager. A concrete example: our agent can read both front-end and back-end errors — server logs and the dev console — something a developer would normally have to go check manually, and by the time this airs it'll also be able to search the web and implement what it finds. It's really about shortening feedback loops and making for a better developer experience, even for users we're not explicitly framing that way, since a lot of our audience is less technical and we don't want to lead with jargon like debugging or console logs.
If you had one takeaway for DevTools people listening, what would it be?
Reevaluate what's actually possible right now. There are things I couldn't automate a year ago, or even three or five months ago — partly because I didn't have time to sit down and understand the topic well enough, and partly because the dev tools and infrastructure weren't good enough yet to support it. Give yourself an hour or two to actually experiment: sit down with Claude, sit down with Replit, and try to get to a solution. The difficulty of doing things is falling fast, and things that felt out of reach recently are becoming possible constantly — that's the exciting part, and it's only going to keep accelerating.
Links & references
- Replit ↗
- Replit Agent documentation ↗
- Diátaxis documentation framework ↗
- Matt's YouTube ↗
- Matt's X ↗
- Replit's YouTube ↗
This episode is brought to you by WorkOS.