We're just like, are we gonna do this again? And then, you know, you you kind of like see the next five or ten years of your life.

I'm joined today by Sam from Mastra, who you may also know as the founder of Gatsby. People want to

build agents, they just don't know how. So in AI world, being able to see what your software is doing is 10 times as important. On an average week, we hand out about 1,500 copies of this book. Really? We we as humans are experiential people.

Like, we learn by playing around and trying things. And the more iterations of something we do, the faster we learn.

You went to Japan recently or something.

We we did. I personally did did not go to Japan but Abi and and Tony from our team did.

That's so cool. And apparent like he was saying that you were that a lot of people were using Maestro in Japan.

Yeah. In Japan for for whatever reason, both, like, tons of TypeScript user folks and loving to tinker with the innovative stuff, we just had, like, sort of it's, like, viral take off. Actually, one of the main reasons that people in Japan sort of adopted Maestro so quickly was that dev Twitter still sort of works in in Japan in a way that it hasn't since Elon bought the platform in in The US. Because the Japanese devs are, like, kind of agnostic to American politics, and they're sort of like, whatever. Like, we're hanging out here.

We'll we'll keep hanging out here. And so you get these sort of you know, people start, you know, just talking about the tools that they like using, and then people start, you know, retweeting that. And then they also have a a lot of, like, essentially similar sites to, like, dev.2 where, like, folks will, like, post articles they write. And so a lot of those type of articles that started going viral on on on on on Japanese Twitter.

Do you know what? It's called the version of dev.co in case anyone's like,

I need to try like a if you search like qlika, and the like .dev is the is the URL.

Okay. Super cool.

But but so we have about like 20% of our users.

20% of your users are in Japan.

20% of our users are in Japan. And and so we we've internationalized the docs into Japanese. I mean, it's the only language we've they're in other than English. Because it was like, hey. Like, I mean, we're we're not gonna sort of, like, you know, do this for any other language.

But if 20% of your users are are in in Japan, you should probably have your your docs in Japanese.

That's so cool. Do think people are missing a trick on, like, the Japanese wildcat for DevTools?

I I think that the Japanese market loves, like, AI stuff in particular, Like very particular forms of of AI stuff. And then if you happen to go viral in Japan, then, you know, just get ready to go fly to Munich in Tokyo where you kind of get to these local celebrities. Really?

Yeah. Super cool. Yeah. Yeah. I actually I spent a little bit of time in Japan and I was going to like a lot of meetups and then and then they have like is they're actually like really big.

I I think at that time, this was a few years ago. At the time it was, I would say, bigger than way more AI events than in London when I was there. Yeah. I don't know if it's still the case.

I mean, I think there's the country has this sort of robotics is also very big in Japan. It might but the the country is very dedicated to kind of and I think, like, the like, culturally, AI is, like, an extension of that as well. You know, I think right what what, you know, what what matters now a lot for AI is just kind of interest because there's a whole huge, like, bulk of people who are trying to become AI engineers and, like, working on sort of small AI engineering projects that could become big AI engineering projects. And, you know, the the relative, you know, driving factor there is is enthusiasm.

Because it's also new. And

Because it's also new because you have to sort of like think about nobody's in very rare cases, people are telling you, hey. We should automate this with AI. It's more like, you know, if anything, maybe, like, the CEO in in the company is like, AI is really important. Please go do an AI thing. But it's really up to you as an engineer to figure out what is the AI thing you want to do and then go do it.

Yeah. That's a good point, actually. So I suppose SF became the hub because a lot of people are like had that mentality.

I mean, obviously, the startup community plays a large role too. Half of our YC batch a few months back was various types of these type like, vertical agent companies, though AI for various industries. And and so, like, we know what does that you know, why see what do I see and what does the venture ecosystem in general do it brings all these people to to SF Bay Area and kind of concentrates them there. And then there's a local meetup scene that, you know, any any night that if you're traveling in SF, you could visit, you know, two to three AI meetups. There's a group called generative AI SF.

If you, like, sort of search for that, you see a bunch of Luma events. And and I think it's, like, you know, it's kind of a self it ends up being a very self sort of perpetuating the ecosystem. You also have, like, all these DevTool AI companies and others that are kind of, like, also want to posture this. They hold the meetups and host the meetups and then, you know, encourage the all the conversation. Right?

So

Yeah. True.

Scaling DevTools is sponsored by WorkOS. If things start going well, some of your customers are gonna start asking for enterprise features. Things like audit trails, SSO, SCIM provisioning, role based access control. These things are hard to build and you could get stuck spending all your time doing that instead of actually making a great dev tool. That's why WorkOS exists.

They help you with all of those enterprise features, and they're trusted by OpenAI, Vercel, and Perplexity. And for user management, the first million monthly active users are completely free. Let's hear from Utpal from digger.dev, a dev tool using WorkOS.

How it's designed is that you can start as early as day zero. But for us, it wasn't day zero. It was closer to when we first started monetizing because we didn't have a sign up at all. People could just anonymously use our tool. So it was a little later.

It coincided with when we wanted to start monetizing and, like, we needed a nice enterprise feature set. If you're open source and you're doing enterprise first, the minute you think about monetization is when you should think about Work OS. To be honest, if we do that again, I think we'd think about that on day zero, to be honest, because, like, should have done it on day zero ideally. Anonymous usage should be permitted, but you should know who's using your tool. It should be optional, 100%.

It should be opt in, 100%. But it'd be great to have auth from day zero. You don't necessarily think about these enterprise features, but they still lead revenue. And it kinda is a no brainer in that sense. So, yeah, I highly recommend.

And actually so Sam, we so I discovered Mastro at at Swigs's World Fair. And it was at the there was a Work OS workshop. And I went along, obviously, OS sponsored podcast. So I'm absolutely gonna go along anyway. But it wasn't it was genuinely such a good workshop.

But it it wasn't really about Work OS. They they just built something very cool with Maestro. It was I think it was like a it was like a GIF. Not GIF. Like a meme generator.

Right? It was like, could write an agent

Is super button there?

Yeah. It was so it was very fun. The people were like feel a of people were like complaining about the WiFi and stuff like that. It was very like lots of conference memes. And it it was really a cool experience, Mastro.

And like the the developer experience is very very awesome. And he written this book as well about agents and building agents not just with Mastro but in general. And so that's that was a very cool way to to come across you and come across across Maestro. But maybe you can kinda share a bit more about about Maestro.

Yeah. I mean, so what I will say about that about that workshop was it was incredibly surreal to build the thing and then show up and and magically, some somebody else is giving a workshop about the thing that you have built and teaching everybody else about it, I I think there's something you you know, I've been we previously, I was the the founder of Gatsby. And so, like, you we we sort of learn to recognize, like, markers along the journey, and that is, like, a very big one and a very humbling and and flattering one. And then second, it was, like, as, you know, Nick Nick and Zach put together, like, a great demo and a great presentation. And and and what happened in in Work OS was was just that somebody demoed it to the entire company.

And then all of a sudden all of a sudden, like, five or six different groups and, like, 10 different developers within the within the company started using it. So and and, like, just organically to solve, you know, various problems, which is also very cool. Maestro is a TypeScript framework for for building AI agents. And so people use it to, you know, hey. I'm gonna you know, for for tasks that, like, might be something locally.

Like, hey. I'm gonna, like, golden agent that, you know, creates Jira tickets for me with all the context from, you know, GitHub issues or, you know, whatever. Or people use it to build to to, like, ship, you know, agent features inside their SaaS app. Or people use it to or people use it to build startups that are building agents in a particular industry, let's say, or a coding agent or, you know, an automating a bunch of work. So Maestro has, you know, a number of different AI agent primitives like agents, workflows, rag, tool tools, evals.

And so we see ourselves as basically a complete framework for building agents in in TypeScript.

Yeah. I've been trying dabbling with agent stuff recently and it's like, there's so many things that you don't think about that I think are that you've tried to make it easier. And I I felt like, you know, in the book you talk about stuff like, you know, how to like update people on what's on on the progress and all these sorts of things that you're not really thinking about that are, like, quite tricky to do. And and it's quite nice that you've got, an opinion on all these things and, you know, just like you're you're thinking about them and trying to make it easier for people to do that sort of like common thing.

Back back last year, we we were actually, you know, starting learning about, you know, agents ourselves. We were building sort of the AI and the agency parts of a of a CRM, and we were kind of going through this, you know, learning journey of, like, what are all the different things you have to do? And and I think, like, what we discovered was that there we were just kind of learning it all ourselves and and using a bunch of different libraries. And we were like, hey. Why don't we just build the framework that makes the the journey really easy?

Did you have, like, a conversation, like, a knowing look of, like, building another framework?

Yeah. Is it I I mean, we we we just kind of, like, you you look at each other and and, like, my you know, we're just like, are we are we gonna do this? Because we we know we can do a good job doing this, but, like, is this what we're gonna do? Yeah. I guess this is what we're gonna do.

Let's, like, let's

go do it. Oh, what what do you think you learned from Gatsby that you're bringing to Maestro?

A lot of the concerns are very similar. So you have to think very deeply about, you know, bundling and and tree shaking in, you know, in various instances. For example, we have, like, a local dev server that people really like using that sort of shows you your agents and your workflows and your other kind of primitives locally and lets you kind of experiment and play around with them. And so, like, we had to do a ton of bundling stuff to get that to work right. You know, just like the you know, knowing the nuances of of CJS and the ESM, but, you know, all all all this there there's a there's a sort of, like, this sort of this kind of, like, niche esoteric knowledge.

API design is I think the other thing that is, like, you know, develop just under a ton of nuance around a developer experience, designing the right APIs, the nouns and verbs that you use to make a framework to make the framework. You know, we we realize, like, we once you ship these right. Once you ship a noun or a verb, you kinda in some ways, you're you're pretty stuck with it.

So, like, that'll be like a what workflows. You've you've got like workflows and agents and tools and

Yeah. And that and that would be the noun and and then, you know, your workflow has a a steps and then the different steps have, you know, there's different ways of dot parallel or dot branch or or dot then. And, like, the we have this sort of fluent syntax for chaining it together. Like, these are these are kind of things. Like, decisions, once you make these kind of decisions, you're you know, it's it's it's very difficult to change them.

Yeah.

Yeah. I I think, like so Maestro started by myself and my cofounder Shane and and Abi who were who were at Gatsby as well. But then, like, we sort of have brought over a number of a number of the most of, like, the engineers from Gatsby have kind of joined as well. Ward Ward Peters, Tony Covenan, Tyler Barnes, who are kind of responsible for, like, build building and and maintaining the the framework over time. And I think, like, we all just have a, like, like, very shared sense of, hey.

Here's how you interact with community the community, you know, and and sort of just whether it's, like, engaging in issues in in Discord, etcetera. And there's a sense of, like, the, you know, really high quality documentation and what that looks like. So I think, like, in in many ways, like, we're building an an AI agent framework. And, you know, some some things are particular to the AI agent bit and some bits are are are particular to the the TypeScript framework bit. And we're we're we're finding that a lot a lot of the things, maybe most of the things carry over.

Yeah. Yeah. And I don't know. Do you wanna is that like a kind of was it a question of like whether this makes sense in TypeScript or like, you know, because I guess like there is like some I don't know what the are more people building stuff with TypeScript than Python in agents? Or is it like still like Python's kinda dominating that?

Or

The the interesting thing has been that sort of, you know, AI development in general has has been in the middle of a shift. The two years ago, people doing AI stuff, it was primarily, like, machine learning, you know, these sort of matrix multiplication type things. And so, you know, the ecosystem started in Python. To do agent stuff, you don't need to do any of that. You're you're just you're building on top of the OpenAI and the Cloud, you know, Gemini APIs.

Right? You're you're managing context windows, the context that you're feeding into the LLMs and and sort of like, you know, doing interesting things with the responses. But it's it's all API driven. But but for legacy reasons, when we kind of came on the scene, the ecosystem was very biased towards Python. And so there weren't really good tools in in TypeScript.

As their tooling in TypeScript has gotten better and and so, like, that's that that's us. That's, you know, other tools like stagehand from Browser based. Is from easy apps from browser based. Yep. Is the tooling is the the shift has changed.

And, actually, a good like, a good way to see that or a good metric is the proportion of startups from each y c batch that are building their agents in Python versus TypeScript. And so Mhmm. We were in the winter batch, so it was from January to March. And, you know, we we knew most of the startups, and and and it was about, you know, 30, you know, 35% of the agent startups were building in in TypeScript. But then we we talked to a couple companies in the next batch that were also very kind of, like, aware of of of, like, all the companies in in their batch building agents.

You know? And this was, like, only three months later. And they said, oh, yeah. It's, 55, 60% TypeScript. And we were just, like, mind mind blown.

Like, what what you know? Because because, like, the the there's so many new people coming into the into the the field and like those folks are wanting to use TypeScript.

That's very cool. That's very very cool. It makes sense. I mean, as you put it like, you don't need to like do matrix multiplication and stuff like that, then why why not just use the tools that you're building for the web? Like same reason, I guess, for other reasons why JavaScript won out, I guess, or dominated.

Yeah.

Firefly Never never bet against JavaScript. I always bet on JavaScript.

Yeah. And so could you also tell us about the book that you've written? Which is very good, by the way. People should should go download it. It's free.

Oh. Yeah. No. So so this is principles of AI agents. I think, you know, we we were having these conversations.

We're bringing in a lot of folks who are building with agents into our our office and just kinda doing these, like, whiteboarding sessions, whereas it was kind of like, describe what you're building and, like, maybe, you know, we'll just, like, walk through it together. And and what we realized was that, like, there's a lot of knowledge that goes into how you design it and just, like, kind of conceptual around, like, what are these different concepts? Why would you use a workflow in one case and then agent in in another case? What is a rag pipeline? What are the different steps?

Like, what are evals? Evals are just test. Anyway, so there are all these kind of topics. And and, like, conceptually, like, was helpful for people to fill in the missing sort of pieces as well as like and so we we realized that we wanted to just kind of like encapsulate and and encode a lot of the knowledge in a way that could be easily shared. I think, you know, like, I was actually listening to a podcast about the Michelin as a company.

And, like, in in in 1900, the like, when there were, like, only a few 100 cars in France, like, these, you know, these Michelin brothers with their tire company wrote like a guide of like restaurants. Like it didn't, you know like it just kind of showed people what to do with their cars and and like where to go or whatever. And and, you know, maybe that this is kind of the the equivalent for for the agent, you know, things like how do you build it, what are the main concepts that you need to know. And and we've we've sort of started we we've we've start we've started just going to hackathons and and, you know, AI engineer and just handing out copies of of books, and we've gotten great kind of, like, reception feedback. And people have said, hey.

This is, like, really useful as I've been, like, getting into agents and understanding what could we even build with agents inside my company, or what could I build as a fun project with agents? And what are all these different things that I keep hearing about, and how how do they fit? Like, what is the knowledge hierarchy that they kind of fit into?

Yeah. It it it is really good. And it's it's very cool, but it's like physical. Don't know. It's like, definitely stands out like having just like a physical book Yeah.

On something which is so up to date. Because I guess like that's the issue of books is like often it's not like super up to date. But your book is like extremely up to date with like tool recommendations and stuff like that, like dev tools to use and things like that, even though it's probably changing every few months

or faster. I mean, the the we we we started writing in February, published the first version in March, and then we published an updated second edition in in May. I think we we we try to keep this thing up to date because it obviously is is you know? So a lot of the concepts are sort of what what happens is that the knowledge is kind of accretive in the sense of, you know so so I'll give an example. Like, something that kind of happened recently is that is that agents got much better at, like, sort of, like, answering questions using, like, web web search tools if you, you know, you if you see deep research and or or, like, Claude has an equivalent.

You you kind of can see this as as well. Right? But but so, like, this is if you, like, do this when you're building an agent inside your company for some you know, for for product use case, you would call this, like, for retrieving knowledge, would call this like an agentic rag, which is kind of a different way of retrieving knowledge using AI tooling than the original sort of rag paradigm. So you sort we sort of already had a rag chapter and now we have, like, an argentic rag section within the, you know, the the rag chapter.

So it's not like a whole, like, wholesale, like, change the whole it's like a yeah. Yeah. That that makes sense. And I guess like your your views on things and what what's important and what isn't will change. But hopefully, the broad topics will be quite similar.

Yeah. They they built like the the the the layers of knowledge, you know, like any engineering domain build on each other. So, you know, if you are a full stack developer and then you start getting into, you know, data engineering, let's say, there's gonna be a set of topics that you kind of learn over time and you become a better end data engineer over time. Like, that that field sort of emerged, you know, five, ten years ago. And and AI engineering as a field is is emerging now.

But in the same way, like, the the the, you know, the the kind of topics are are, like, very then and knowledge sets are sort of built on top of each other. I I think that's, like, actually one of our big lessons to to engineers who are are getting into AI engineering is that it's is that, like some some people, like, find the terminology that, like, can be used in the in the jargon in the field, like, off putting or just, like, frightening or whatever. And we're like, look. It's you know, a lot of the the concepts that you know will translate translate over. When someone says evals, they just kinda mean tests.

And your intuition around writing tests will translate over to this new world. It's not a different thing. It's just an extension of your your existing knowledge. And and and by the way, I I I don't think that was always true. I think, like, two years ago, you know, the with the types of ML that and and that you'd have to do, think that was, like, a very different kind of field.

But I think now it's, you know, agent AI engineering and building agents is is really just a or or an assistance and other, you know, related things is really just an extension of of what we do as as a as software engineers. And

yeah. Because it's a you're basically just the AI is the API calls that you're making. So Yes. True. Yeah.

Yeah. That that makes a lot of sense. How do you see Maestro, like, kind of evolving? Like, you're you're building, like, a hosted version at some point. Right?

I've I've seen it kinda mentioned.

Yeah. We so so we recently launched Maestro Cloud into into public beta, which is, you know, basically a hosted hosting. So you it's like a runtime a serverless runtime from, you know, from for for your agents, you know, syncs with your GitHub commits, and you you push in, you know, a new new commit, and you can deploy a a new version of your of your agents that's, you know, hosting. Right? And then sort of, like, a little bit of observability.

So that's tracing seeing, you know, tracing in in evals. And I'm gonna sort of, like, digress a little bit before I come back, but it it turns out that in AI world, like, being able to see what your your software is doing is, you know, 10 times as important as it is in in normal software engineering because LMs are nondeterministic, and so they just diverge in more interesting ways than the normal engineering does. And so this kind of, like, notion of observability becomes even even more important than than it is normally. So so I mean, the the so I think, like, from a we think that, like we see the journey on as basically, you know, building the best TypeScript framework for building agents and then building kind of the dedicated cloud, you know, hosting and and observability, you know, platform from that. Obviously, our our lessons are very informed by, you know, having built sort of Gatsby Cloud as a serverless hosting platform for for Gatsby sites and, you know, scaling it to a few million ARR and and then also obviously seeing Netlify and Vercel and, you know, their other platforms.

Again, like I think yeah.

Oh, that's super cool. Yeah.

This is actually, you know, another place where where there's a lot of sort of carryover in that from from kind of Gatsby's and that, you know, building a a a pass, you know, platform to run your user's code is sort of substantively different from an infrastructure standpoint than building, you know, a SaaS. You you know, we have this, you know, have this Kubernetes setup on on Google Cloud that we did last time, and we're, you know, doing it again with the same you know, Avi, you know, my co cofounder and and and and Tony Covenant, who is a built built the first version of Gatsby Cloud, and now they're, you know, building they they built, you know, cluster cloud. So

Yeah. You

know? And it's it's a you we're we're just like, are we are we gonna do this again? And then, you know, you you kind of, like, see the next, like, five or ten years of your life because you know exactly what the journey is. Yeah. Like with with a lot higher precision than you usually do in starting a company when you've when you've started the previous version of the company and run it for several years.

Yeah. That's very cool though. But it's like it's like the mechanics are the same, but very different, you know, very different area, very different products. Yeah. But like same kind of like, I guess Yeah.

Like you're you're on the you've been on the journey to to mastery.

Yeah. Yeah. I mean, we you know, okay. Like, you know, a lot of times it's like, you know, well, what parts of our old GKE setup did we, you know, not like or, you know, gonna just get what about our, you know, infra analytics pipeline to, you know, or did we not, you know, like or not like, you know, okay. Like, it's sort of like if you if you got to build a similar version of the the old thing again, you you know, you get to walk through your engineering tree of decisions again.

And and so you have some knowledge of of what's in front of you, which is, you know, both both fun and and good.

But Yeah. It's that's very very cool. And it seems like Mastro has been taking off a lot. I mean, established big in Japan now as well and taken over AI Engineer's welfare. What what have you been doing for marketing and stuff like that?

Like putting putting the word out?

I think like a lot of it so there's a variety of different things. So number one, I think that your best form of marketing is always your docs and your change logs. And so you spend a ton of time making sure those are really good and and complete.

So this is a dumb question, Sam, but how does the change log like, drive the growth? Is it like the people start talking? Is it you mean like because people see that it's growing or like

People seem the regular cadence of feature releases as a sign that the project a sign of life of a project from, oh, this thing is completely dead to, yeah, like it's on maintenance mode to Yeah. Like They're pushing. Oh yeah, you know, the team's shipping some stuff to like, the team is shipping stuff. This is exciting. This could be like the new thing.

And like Yeah. Yeah. You know, people see the sort of like the volume of, you know, things that we're pushing out, you know, day to day, week to week is like, yeah, this is like the like, you know, that was like a feature they did last week and this is a feature they're doing this week. And you know, so so I think, like, that's and then we push them out through social. You know, we we we push them out through social.

We also have, you know, published things on on GitHub when people come to browse the the project, the sort of which is more of like a PR level view. These are the 35 PRs we shipped this week, and here's the links to each one in a short description. But I think these are you know and and like people people gravitate, rightly so, to projects that they think are not only you know, they look at the feature set, but they also look at the velocity because it's, you know, it's it's position and it's also slope of the of the project. And and so, you know, you can you can we inherently as developers judge the slope of a of a project by by this kind of like change log type stuff.

Yeah. That that makes sense. And sorry I interrupted you though. You were saying that primary like the docs and changelog are the driving force.

The so the docs and the the changelog are like I I think like then we we spend a lot we we've been publishing, you know, two two or three sort of stories of, you know, Maestro users every week, just kind of like putting them on their blogs, of like pushing them out pushing them out through through social media, which is again, it's sort of like the the book. It's like people want to do this and they just the best social proof they the more social proof they have that like, hey, this could work for me. That's kind of like very helpful and useful for them. So we we, you know, we we we probably like we we hand out about and and I'm gonna say this number and you're gonna laugh. But on an average week, we hand out about 1,500 copies of this book at at various events and and and conferences.

Because, you know, like, a lot of times, people are just people want to build agents. They just don't know how. And so if they see a book called Principles of Filting AI Agents, they pick it up, and they they start reading. And they're like, oh, I guess this is how I I do it. And so so so that that, you know, that that drives growth sort of significantly.

And then people will will then, like, you know, post about post about the their experiences with building Moxtra or reading the book on social media and will will repost. But I I think, like, like, like, the other thing is, like, I I think it's you know, you can point to, here are the external things we do. But, like, the core growth of any developer tool is people telling their is well, there's two things. There's one is people telling their friends, and then then, like, we we as developers are very analytical about how we make. We're like, well, I need a tool that does x thing in this you know, I need, a, you know, I need a bundler or I need a, you know, I need a project for my to set up my mono repo or, like, I you know, I need a TypeScript framework to, you know, build agents.

And they then they kind of analytically are like, what is the right tool for the job? And so if you if you're kind of showing evidence that you are the, you know, best tool for the job and and then, like, they kind of gravitate towards you naturally. So those are like you know, I think like the things we spend a lot of time thinking about both of those different pieces. You know, how do you get people to tell your friends? Well, you need to sort of create something that people love.

One of the things that people love the most is our our dev server because it gives you this like real time sort of like feedback on what you're you're doing. And so we just spend a lot of time polishing it and making it great. And every time someone's like, hey, I wish that you could show me this information in this way, we like add some way of of of doing of of doing that. And it's a really cool sort of you you end up you you know, and just like because we've like obsessed about that, then like, people can tell that we've obsessed about that. And then they're like, this is masterless.

That server is really cool. This is like it was just so neat. And then they those are all their friends and you know, the cycle goes on.

I guess it's very it's very visual, your dev server, like the dashboard. And then also it's like the first thing you do kind of as well. Right? Like so it feels like the fact that it's really good and it's like visual and it's like right up front. Like, that's the that was the first thing I saw when I, you know, set up, you know, to to follow the getting started guide.

It's like, check out the dashboard. Oh, wow. Cool. Like, I can see all these things that I wouldn't otherwise have known how to see and interact with it.

We we as humans are experiential people. Like, we learn by playing around and trying things. And the more iterations of something we do, the faster we learn. Right? And that's why little children, you know, do try so many random things because they're they have every couple of kids.

Like, they're just experimenting about the world, and I think, like, we wanted to create something. Like, oh, I will okay. I just, like, got an agent out of the box, and I could poke and play around with it. And then the when the agent does changes, and I can, you know, test out workflows. And it's it's just very, like it's very experiential, and it, like, you with that, like, learning loop.

But it also helps you with the, okay. Cool. I've learned how this all works. But it also helps you with the, hey. I'm like just iterating on this and making it better.

Because it's really the same kind of thing, just the different levels of fidelity or different levels of, you know, knowledge.

Yeah. Yeah. Yeah. It's it's very cool. Experiential.

Make it. Yeah. Playful. Yeah. Very cool.

I have to ask you about the books. I mean, that sounds like 1,500 a week. Yes. Like logistically, how are you how are you pulling this off?

I mean, we there's a lot of meetups in in SF and conferences in SF, you know, we also travel to other places. The book is on Amazon. We published it through Amazon. So, you know, we are just honestly piggybacking on a very talented logistics company that will ship the book to wherever we need it to, which is usually like to, you know, our

Oh, okay. Okay. Okay. So the so a lot of these are like just like people order it through the website and stuff. Because I was thinking or you just hand out at the meetups, like, 15 No.

We do. We we we will order a box of, you know we we will make an order of 500 books and then we will, you know, show up in Atlanta and just distribute them at at render. Right? Or and we'll you you sort of like you just you go around the conference and you kind of leave them on the tables, then they sort of like vanish into you people's like, when I was a kid, like, when I was a kid and my mom wanted me to read something, she would literally just get a book from the library and she would, like, put it on the ground. And I would, like, go sit on

the ground

and read the book. It's it's like nerds typing. Right? Like, you're you're

like Yeah.

You're you're you're you're you just, like, are you really interested in really you get really interested in stuff. And so if you, you know, if you so we go around conferences and we we put them we put it around conferences and people pick it up and people read the book and near events and meetups the same way.

Very cool. Very cool. So Bezos is a double win on you if like if it's like AWSS. I don't know if you'll you said you're GCPA ing me, but Or

you know So maybe not

too. He's lost out on that. But

Yeah. Yeah.

Let's make it up for our old books.

Yes.

Yeah. One thing that I wanted to ask you just because I'm sure like some people are kind of buzzword, but you've written about MCP in your book. Yes. And as a kind of like stepping back, like should all DevTools be building an MCP server at this point? Or

I mean, I think the MCPs are great for like, a number of you know? I I I think, like, every every time when you're building an a API, right, a publicly accessible a publicly accessible API, right, You should you should consider, like like, hey. Should I also be building an MCP server here? Right? You see see if the publicly accessible API is for normal traffic, maybe the m c the MCP server could be for agentic traffic.

I think it's a really I mean, so so for context, for you know, MCP is a new protocol that Anthropic has kind of put out about. Essentially, think about it as untrusted tool used by agents. Let's say I have an agent that wants to, you know, access somebody else's information. Like like and and you need to do this for any sort of like, hey. I need to, you know, browse the web or I need to, you know, access data in some location that maybe it's in Google Drive or, you know, whatever that I don't necessarily control, but, like, that agent needs the data in that location to be able to answer a question well.

And so MCP sort of went viral as a as a as a protocol. I mean, I see it as kind of like this agentic successor to to open API. Right? And that, like, it's it's a it's a good sort of, like, framework for how to act so if you're building with agents, I would say, like you know? And and there's also really cool things you can do with it.

We have a, like, a MCP server for our docs. And what that means is that, you know, if you're using Cursor or WinSurf or Versus Code, you can install the the master MCP server, and it will have up to date, like, you know, it will it has tools to grab up to date versions of all the code and and docs and and reason about them and throw them into context for your your cursor, windsurf, you know, agent.

I saw you actually have that in your installation process as well. Right? It's very cool. How you just say, like, do you want to make your do you want to make your agent an expert on Astra? Mhmm.

And then it

Yeah. It

was, like, really cool.

Because it it turns out that, you know, like, while reading the docs is is one way of, you know, learning to build with the framework. Also, if your agent that's writing most of your code for you or at least writing a first draft of most of your code for you anyway locally on your machine can read the stocks as well. It can, you know, generate higher fidelity, you know, first drafts of the of, you know, whatever you're trying to build. So I think there are also, like, lots of these kinds of, like, use cases. There are that that are, like, neat or interesting or cool for for folks to explore that of, like, you know, maybe they could make build an MCP server.

You know, like like, if if you if you're trying to analyze some dataset, for example, like, some medium sized dataset, maybe it's, like, your corpus of blog posts or whatever, like, consider just, like, writing an MCP server that, like, you know, you gets has the tools that you would give a human analyst. You know? Get all posts, get posts by category, get specific posts or whatever. And then, like, you know, search search across posts, you know, and then, like, make some MCP server that wraps that that wraps, you know, these kind of tools. Right?

As MCP servers are containing for tools. And then then install it on your local on your local Chris or Windsurf and and start asking it questions. So one of our investors, Elena Goyal, did this for her blog, and then she's you you this is a great demo where she's like, you know, like, you you it's like a recording of her, like, asking, you know, asking like, typing in, you know, the Windsurf agent and asking her questions about herself and it's like retrieving information about her from her blog.

Oh, yeah. I saw that from your book actually. Right? It was like and they she asked like, what's your fave who's her favorite portfolio company? And there's like, no comment or something.

Right?

Yeah. Yeah. He gave some polite some very polite response there. It's like, well, these are her favorite restaurants and then like, gave a very informative response there and just give like a polite non reply on on that particular question. She loves the Mollie Cooley.

Yeah.

That's very cool. Yeah. She was like a big theme in your book actually. Feel she got like a bot. I think even like when she wasn't mentioned it was like our investor and I was like, I know for a fact that that is a lot of

Yeah. I mean, unlike it's just very unusual in that. She's just, like, the for an investor that, like, she loves playing around with all sorts of tools. And so She's been great. She's given us a number of, like, useful, like, and helpful, like, feedback.

Though there's one weekend in particular where we where we you know, Shane, my my cofounder, and and her were sort of, like, you know, basically ended up debugging the differences in how the cursor and Windsurf agents were implementing the MCP spec. Like, there were, like, these subtle and this is, like, okay. We, like, shipped sort of the correct version of our MCP server, but it turns out that the Cursor and the Windsurf agents have implemented the spike slightly differently. And so we actually needed it. It was it was a it was an interesting and fun kit.

It's added value. Added piece of value. Yeah. Sam, this was really awesome. I I wanted to ask if you got any kind of shout outs, say, if you wanna send people to to get the book or or

Yeah. I mean, shout out. So you can get the book. If you want a digital copy for free, you can go to masra.ai/book and grab a digital copy. You can grab a physical copy in on Amazon.

If you want to install and use master, it's n p m create master at latest, and it'll walk you through that that, like, the the setup guide that that you kind of mentioned. So from the command line, I think, like, you know, shout out to I think shout out to to all to both, you know, the master team for, like I mean, just we've run a great group. Like, I I think, like, I I get the privilege of coming here and and chatting with you, but I'm really just representing the word that, you know, a team of really incredible human beings who have the privilege of, like, getting to work with for the second time have have done. So shout out to, you know, all the the members of the Hofstra team.

Shout out, master team. Yeah. Well, thank you very much, Sam, and thanks everyone for listening.

Thanks for for having me on, Jack.