How I AI - How Block’s custom AI agent supercharges every team, from sales to data to engineering | Jackie Brosamer & Brad Axen

Episode Date: July 28, 2025

VP of engineering Jackie Brosamer and principal engineer Brad Axen join me to demo Goose, Block’s open-source AI agent that runs locally, plugs into your existing tools through model context protocol (MCP) servers, and peels away the rote parts of work so people can focus on insight and impact.This episode is packed with in-depth demos: starting with a messy farm-stand sales CSV, Goose analyzes the data, builds visualizations, and generates a shareable HTML report. We then spin up an MCP that lets Goose talk to Square’s dashboard for inventory management, vibe code an email MCP that can send payment links automatically, and unpack how environment setup, debugging, and tool orchestration get handled behind the scenes.What you’ll learn:A practical, repeatable workflow for turning any working script or function into a custom MCP—and exposing it to natural-language controlHow to transform messy CSVs into visualizations, HTML reports, and actionable business insights without needing a data science backgroundWays to hook Goose into live business systems (e.g. Square inventory, payments) so analysis flows directly into operational actionThe thinking behind Block’s decision to open-source GooseLessons from Block’s bottom-up meets top-down adoption modelWhy organizational transformation, not just picking the right LLM, will separate AI winners from laggards over the next few yearsHow to scale an internal MCP catalogThe organizational transformation required to fully leverage AI capabilities—Brought to you by:CodeRabbit—Cut code review time and bugs in half. Instantly.Lenny’s List—Hands-on AI education curated by Lenny and Claire—Where to find Jackie Brosamer:LinkedIn: https://www.linkedin.com/in/jbrosamer/—Where to find Brad Axen:LinkedIn: https://www.linkedin.com/in/bradleyaxen/—Where to find Claire Vo:ChatPRD: https://www.chatprd.ai/Website: https://clairevo.com/LinkedIn: https://www.linkedin.com/in/clairevo/X: https://x.com/clairevo—In this episode, we cover:(00:00) Introduction to Goose and its data analysis capabilities(02:27) How Block embraced AI across the organization(04:48) What Goose is and why Block open-sourced it(07:45) Demo: Analyzing farm-stand sales data with Goose(12:18) Creating shareable HTML reports from data analysis(14:15) Model context protocols (MCPs) that Goose uses(18:56) Demo: Using Square MCP to create a product catalog(23:35) Creating payment links from analyzed data(26:30) Demo: Building a custom email MCP(31:18) Testing the new email MCP with Goose(36:09) Debugging and fixing MCP code errors(38:44) Connecting workflows: sending payment links via email(41:30) Lightning round and final thoughts—Tools referenced:• Goose: https://block.github.io/goose/• Pandas: https://pandas.pydata.org/• Plotly: https://plotly.com/• Python: https://www.python.org/• ChatGPT: https://chat.openai.com/• Claude: https://claude.ai/• Cursor: https://www.cursor.com/• Mailgun: https://www.mailgun.com/—Other references:• Block: https://block.com/• Model context protocol (MCP): https://www.anthropic.com/news/model-context-protocol• GitHub: https://github.com/—Production and marketing by https://penname.co/. For inquiries about sponsoring the podcast, email [e-mail].

Goose is our AI agent. We kind of designed it to be really agnostic. You tell it what you needed to do by connecting it to different capabilities, and they can just solve any problem. What I'm going to do is take that CSV that Jackie was working with. I'm just going to pop it over, and I'm going to say, can you read through this data? Use it to create items in my square dashboard. So we took a CSV that we didn't even have to look at, that probably had data that was not in this exact format, and you created a product catalog. What a lot of my work and what my team is trying to do

is not just make it faster for someone like me or on my team that's data expertise to go through this, but to allow our finance team, our sales team, anyone in the company to be able to dig in and self-serve a lot of this data rather than having to ask an expert and wait for that to come back. You have been the first person that's told me

it was the salespeople that were begging for it. We hear so much, you know, engineers are leading in and they want to vibe code, they want to do all that. But it's nice to hear that the folks close to customers and close to revenue are also seeing the value. Welcome back to How IAI. I'm Claire Vow, product leader and AI obsessive here on a mission to help you build better with these new tools. Today, we're speaking with Jackie and Brad at Block, who are going to show us how their open source AI agent, Goose, can be used to do everything from vibe data analysis to vibe coding and MCP.

Let's get to it. This episode is brought to you by CodeRabbit. the AI code review platform, transforming how engineering teams ship faster with AI without sacrificing code quality. Quality code reviews are critical, but time consuming. CodeRabbit acts as your AI co-pilot, providing instant code review comments and potential impacts of every poll request, beyond just flagging issues. CodeRabbit provides one-click-fix suggestions and lets you define custom code quality rules using AST grut patterns, catching subtle issues, that traditional static analysis tools might miss.

CodeRabbit brings AI-powered code reviews directly into VS code, cursor, and windsor. CodeRabbit has so far reviewed more than 10 million PRs, been installed on 1 million repositories, and has been used by 70,000 open source projects. Get CodeRabbit free for an entire year at coderabit.aI and use the code, how I, AI. Thanks for being here. having us. Yeah, excited to chat. I am so excited to have you two on Howay AI because I have been so impressed how Block has as a full organization and a large organization at that really embraced AI, it seems everywhere and for everything. And, you know, everybody's worried about AI native

startups, but you're one of the companies that I really think is going there very fast as a much larger organization. So how did you lean in as an org so fast and so broad? I think it's a combination of both like bottoms up and tops down. We saw, you know, a lot of concentrated pockics once chat GPT came out that were really finding a lot of value out of these tools. For example, salespeople, they're like, any tool give it to me, like I'll use it, help me get those leads. And we also saw a lot of engineers, you know, starting to lean in as these tools became really good at coding. And kind of as that started from the bottom, I think a lot of our leadership saw a lot of promise and they've been, you know, trying to make some of those practices a little bit more consistent among groups.

I think that it's really a great time to focus on organizational transformation because to me, that's really what's going to define who ends up the winners in whatever this AI future is, is who not just leans into this technology, but even more than that leans into the organizational transformation, which can a lot of ways be even harder. It's like really easy to get a technology to go exponentially and humans don't go exponentially so well. You have been the first person that's told me it was the salespeople. It was the salespeople that were begging for. We hear so much, you know, engineers are leading in and they want a vibe code and they want to do all that. But it's nice to hear that the folks close to customers and close to revenue are also seeing the value.

Yeah, I think it's, you know, everyone finds like that different first use case that really gets them hooked and one of the things I've been really impressed with with some of our non-developers especially. Like we all kind of know how people use it for vibe coding. And I think that's like well discussed is non-developers are just so creative and so tolerant of like stringing together a bunch of different tools just in a way that you never would have imagined. And that's, I think, really going to be the power is getting these tools closer to the people that actually understand the problems and see what wacky things they come up with with this new technology. Not only are you all leading the way in terms of how organizations, and I totally agree with you, I think the biggest transformation is

cultural and operational. The tech piece is like a lagging piece of it you can figure out. But what I also love about what you're doing is not just how you're embracing it internally, but how you're trying to drive and help others adopt it externally. And so I'm wondering if you tell us a little bit about the adorably named Goose and what Goose is and why you all build it. So Gooses are like AI agent. And And a lot of people have a lot of different definitions for what an AI agent even is now. But I'll say specifically what we mean by that is it's powered by an LLM, but it has a whole collection of tools that it's going to go use to solve problems for you.

And this is like something we're really excited about, in part just because it lets us like build for ourselves and like come up with workflows that work for our company. But like you said, we open sourced it because we think there's a lot of extensibility for everybody and we kind of designed it to be really agnostic. Like you you tell it what you needed to do by connecting it to different capabilities and it can just kind of solve any problem and I think we'll get to show some of that today. And that's all powered by MCP. So that's a big part of this like open source puzzle is how do we make this work no matter what you wanted to do? And MCP really helped us get in like early into helping to establish the protocol and have a way for us to connect it to

anything and people can develop their own MCPs and just make things work. Got it. So you, and I just want to show here, it's up on GitHub, it's open source. It's your agent, your agent framework powered a lot, it seems by MCPs, which I think we're going to talk about later that anybody can can use and extend upon. And what made you all think about open sourcing this? Yeah, open source is, I think, a really big emphasis just as a value for our company. And I think there's two ways we think of it. Number one, this is just like the right thing to do, you know, for the world in some way. But then there's also, we think that there's so much benefit in terms of having this open ecosystem and really seeing these patterns emerge from not just,

you know, the engineers that we have in our company, but from engineers all over the world. And I think we're really seeing that come to fruition with a lot of the growth that we've seen in MCP, you know, and all the MCP servers that are out out there. And it's only been, you know, what, like six, eight months. And then I think we also wanted to really lean into open source in order to take advantage of all these different models that are coming out. A lot of the tools that come bundled, you know, released by these different model providers can only use like that kind of models. And we know that every month a new models coming out, one might be better at coding, one might be better at writing. And we think there's a huge amount of power in being able to

bring together data across all these different NCPs and then also pair that with really any model that you want. Okay. So speaking of data, I think you're going to show us how to do vibe data analysis, which is a first for how I AI. So it's a flow that you use internally. So this is definitely something we're not making it up. You use this exact flow day to day, but we're using some fake data so we can screen share and really get into how it all works. And so. Jackie, what kind of business are we today that we're going to do some data analysis with? Yeah, so I tried to make this as realistic as possible. When I don't work in tech, I actually have a small farm that is not very profitable. And so we're going to be doing some analysis that I've actually

done with my own internal data that kind of shows us how we can look at some business data and make some recommendations. So what we're going to do is go ahead and look at this to try to find this local CSV of data that has July data for my farm stand, which does a bunch of vegetables. And what we've asked is we've asked Goose to use Pandas, which is a Python library, to tell me the items how to have the best revenue and what the busiest days of the weeks are. And then I left a general kind of question at the end, just what are some trends? Because I think it's always interesting when you get a slightly different result as you run this.

So you can see the first thing that we're doing here is it's using RipGrap to find the CSV. We found that RipGrap is actually really effective, even though it's really simple. There's a lot you can get done with it versus something fancier like a semantic search. It identifies this July data at CSV. Then it writes a Python command. It found some error where it didn't have this installed. Oh, no, this is maybe going into cursed virtual environment territory. but it looks like it's going to solve itself.

Sometimes I'll tell it, like, check your Python environment so that it makes sure it sources it, but it decided to go ahead and create a new virtual environment here. One of the things I really like about Goose is I'm terrible at maintaining developer environments. And so it will go through and try to just, like, debug that for me. I mean, how many of us, maybe this is a very niche comment, have been victimized by trying to get the right version of Pandas installed? Because I definitely have. Okay, and just taking a step back, we didn't see the input data,

but is this just like general day by day sales of different products exactly this was just a list of you know item quantity date price and so from that it's able to go through and say okay you know your top generating your top revenue generating items are berries we love berries very high margins and it's got the best days are like Thursdays and so and then I just had this you know bonus question or like what are some key items and trends it'll kind of lead into some stuff later. So cherries have the highest average value. It goes through, you know, what are your mid, high, low items, and then, you know, goes through some weekly patterns. And so without even asking, of course, it comes out with some recommendations, which I really love.

Let's just pause on these recommendations. So what I love about this is you have this ugly file, you probably could have tossed in a spreadsheet or done some of this analysis yourself, and you're doing sums, and you're doing pivot tables, and you're trying to explore this, and you probably have to do different tabs for all these different analyses. And this you can really drop in. And it's maybe a relatively, if you scroll up, it's, you know, relatively simple analysis by day by product. But under the hood, you have this pretty powerful data analysis package. And so I'm presuming you can go from this level of analysis to something much, much deeper. Yeah, that's right. And I, when I'm doing something like this live with something like use, what I'll usually do is kind of, you know,

have it take a high level first pass and then, you know, dive into specific sections and say like, hey, can we go deeper here, you know, use this particular type of method, you know, do a line regression that tries to figure out price elasticity or something like that. Well, and I have to say as a mom of young boys, I do feel like I spend, you know, $1,500 on strawberries every month. So this is really hitting home. Okay, so we can go down. And then I'd love to see, I like the idea of when you're prompting a data analysis to say, give me the data analysis, but also give me some recommendations. And so it's really saying here, you know, high margin items, which is great. Look at why Friday drops because, you know, maybe they would expect that to be a higher day,

weekend sales. And then some, oh, price adjustments. That's really interesting. So what would you do next after this? Yeah. So I don't think that this is like very good to share. And so a lot of times what we see is that people really like to share these kinds of analyses as local webpages, especially people who weren't technical.

just a whole new world is opened up with them to them as they figure out how to use AI to turn some of the stuff into HTML and share it with each other. We see a lot. We have a way for people to use Goose to share these internal static websites. And we see that really popular, especially among, you know, some of our business users that want to kind of create their own dashboards for, you know, whatever. While this is generating, I'm just reflecting on what I've seen so far. So it seems like Goose has sort of a virtual machine in which it can write code and wrestle with. with Python dependencies. Love that for for Goose. It has the ability to spin up hosted HTML and websites. Is that right? Is that what it's kind of doing here? It's all running locally. I think that's

one thing that's like really unique about Goose. And part of the reason, you know, we started that way is that we a lot of our developer environment stuff happens locally. And so that was like a really natural fit. And then we also saw within the open source community that this has actually been really popular with kind of the crowd that really wants to maintain end to end. And control over their workflows. One of the things we've seen with Goose is that some of the most promising use cases are, you know, you tell Goose to do something, you kind of go away and, you know, wait for that task to complete as it does things like, you know, write code error errors.

Oh. And so it popped. It's also cool because it can, it went ahead and pop this up for me. Yep. This is not the prettiest website it's ever made with this fake data. But you can see, you know, it's able to make these bar charts with Plotley. And I could go ahead and dig in and try to refine this if I wanted it to be, you know, a little bit more well designed. Got it.

And, you know, if I was being cheap and easy about sharing it, I would just print as a PDF and shoot it over and it would look nice. When we run this internally, we have an MCP that it connects to where you can just publish it to the company. Got it kind of keep it internally private, but share it with anybody. And what are the MCPs sort of out of the box that if I cloned the good? Goose repo right away, would it have ready to go? And then what are some kind of ones that you've either built or integrated specific to maybe what you do at Block? Yeah. So we have like a collection of different MCPs. Some of them we've taken from the open source community. Some of them we've developed internally because, you know, for example, we have an internal data MCP that I use

every day that helps, you know, sort through where's the revenue for X product. I mean, it knows all that context and we're also working on trying to have external MCPs that, you know, connect to our business products. A lot of what people are using use for and a lot where we saw the first value is just MCPs that connect to all of our key internal system. So Google Drive, probably the one I use the most. Brad mentioned this one we have internally called Block Cell that's really popular because it is a click of a button, you know, you can publish a website and send a URL to anyone in the company. We also have a developer at Stention that does a lot of this like command line and tool stuff. We also have a computer controller one that just did. And that's how it just

opened this Chrome window. Got it. Cool. That's the stuff that bundles with Goose. So when you get it, it's like you're going to have like five really simple like memory developer. Yep. And then when we use it internally, we have like 40 or 50. Yep. And Jackie, how often are you doing an analysis like this in your week using Goose? If I'm doing an analysis, I'm using Goose. The question is more like how many analyses do I do per week? I would say that I probably actually use it more in my day-to-day life for writing, but there's people in my team that, you know, are professional analysts or professional data scientists that are doing this, you know, every day. And I think increasingly what a lot of my work and what my team is trying to do

is not just make it faster for someone like me or on my team as data expertise to go through this, but to allow our finance team, our sales team, you know, anyone in the company to be able to dig in and like self-serve a lot of this data rather than having to ask an expert and wait for that to come back. I think we've all been talking about like the age of self-served data for a long time and it's like kind of gotten better, but we all know that you usually still go ask a data scientist. And so I think it's really interesting to think about how can we, you know, kind of short circuit that and allow everybody to focus less on kind of the rope parts of data analysis and spend more time talking about, you know, the insights and the context and the creative

parts that we all want to spend our days on. Okay. So we've talked about the salespeople, love their AI. The farmers love their AI. But can we please get back to what people can wrap their head around, which is the software engineers? So you have this, you have this great, great data and you want to do something with it. And so you have this data, you know, maybe you want to do something with it to improve your business. Brad, what's the next step in your mind here? Yeah, let me share some on my screen for this.

And we can show a little bit about like how a developer can build MCPs and use that to actually like connect to what someone is doing into like existing products or to like solve new problems. And this is really how I spend my day to day. Like I feel like at this moment in time, I'm just developing MCP select and right. And I'm going to just pause right there because no matter how many times people hear it, they still need an MCP explain to them. So Brad, we're going to take your explanation of what is an MCP. Let me do that by showing you the Patriot and Goose here.

So in this, we have just this huge collection of extensions in Goose. And these are all implemented with model context. So each of these is an MCP server. And what that means is that it has a collection of tools that it exposes to the agent. And it can do a couple other things like provide data resources and all that other stuff. And so it's like a protocol that lets the agent talk to a third party and get things that it can do and data it can read. And most importantly, it's kind of the arms and legs for the model. So this is how the model goes and interacts with the real world and like makes changes in online systems.

Okay, so you're spending all your time building MCPs. It seems like everybody is spending all their time building MCPs. So why don't you walk us through how you're spending your days? Yeah. And so let me start by showing you an existing one before we start building. And so I've turned on a real square MTP server. So this is like from this is available like as an online MCP server that people can use to like run their business. So this is my square dashboard and it's totally empty right now.

This is a kind of an account I set up for the sake of the demo. And we're going to connect that through our agent and actually have it fill things in. So what I'm going to do is take that CSV that Jackie was working with. I'm just going to pop it over. And I'm going to say, can you read through this data and use it to create items in my swery dashboard? And so this is something. that you could do with all kinds of input formats.

Like, for example, I could give it an image and like maybe a picture of a menu or something like that. And so since it's multimodal, it's just going to go figure out that source translated into the API calls that it needs. And I'm thinking as a product builder who has built many import your data products where you're always trying to say, like, it has to be a CSV in this format and these columns with this column name. And what I've just realized this opens up for you is like,

like drop in your arbitrary, messy, whatever data, don't have to reformat it. Let the MCP figure out the translation and then get it into the system. Okay. Yeah. So it doesn't matter what you've got. It's going to figure out, it could be a PDF, whatever. It's going to do its best to like figure out how to make that work. You know, the first thing it does is it just reads it.

Yeah. And it's then going to go, this is actually a really interesting part. I think the Square team did a great job designing this NCP because it has so much. so many different things that you can do on square that it's too much for the model to handle all at once. So what it actually does under the hood here is it says like, oh, if you need to know about catalog, I give you a method to look up what you can do in catalog.

And then under the hood, the LOM is seeing like all of these operations that it can now access. So all that stuff about like creating items that figured out. Okay, and I actually see the, it's like, this is a funny, I normally would just read the CSV. So I'm like laughing a little bit at like how much work it's going through for this. But it is what it is.

I know. I have I have that same experience using like a cursor max mode where I'm like, I just don't know if you need to work to think that hard. Exactly. I do this. But you know, it's not my brain. It's the AI's brain. Yeah. Yeah.

So we'll see. I'll let it cook for a minute here and see how it figures this out. But effectively what's happening is it's just deciding to use. shell commands to grab the prices from that CSV. And then it's going to turn that into an API request and actually send it to the square dashboard and create like a product category and then fill in the items. So we should just be one request away from getting that solved. Got it.

And, you know, instead of sitting here and waiting, Jackie, I'm presuming farmer Jackie would have to go in here and one by one sort of enter, enter this data in. make sure it's right and all that kind of stuff. One of the like really underappreciated things about LLMs is like how much they function is like data duct tape. You know, it's like it doesn't quite fit. You can do something to just kind of make it all mush together. Yeah. And that's where a lot of the day-to-day data work is is trying to match all those things. How IAI is now on Lenny's list with my personal selection of the best AI engineering courses on Maithen. You can spend months, thinking and playing with AI before really integrating it into your workflow or shipping an actual

AI feature. If you want to start building, then these hands-on maven courses are for you. Learn directly from Ishwarya Nuresh Riganti, MIT instructor and AI scientist at AWS, or Sanders Shulov, who has authored research with Open AI, Hugging Face, and Stanford. To pivot into an AI role or successfully lead your company's next AI initiative, visit maven.com slash Lenny to enroll now. Use code Lenny's list for $100 off. That's MAV-E-N.com slash Lenny to get ahead in the AI era and start building. Okay, look at that. So we took a CSV that we didn't even have to look at that probably had data that was not in this exact format. And you created a product cap. out of it.

And this is something like I could follow up with this and say like, okay, grab images or even generate images and like upload them on. Yeah, that's what I was going to ask. Yeah. I want to AI generated Apple image. Exactly. Yeah.

And the, I'll skip it for the sake of time. But like there's so much you can do to just like navigate through all of this. And this is something where like we're watching it live for the sake of the demo. But really the best version of this is like handed the CSB and walk away and come back. and then you're going to just see that this is done. And so that kind of like background task thing is the, I think the best mode for Goose in a lot of ways.

But yeah, it like goes and figures this out. And this is like a real account. So what's cool about this is that this isn't just like a little website showing that I have prices for apples. It's something I can take payments with. So I can now say, okay, let's make a payment link for three pumpkins. And we'll let that run for a little bit too. This is not the topic.

thought I was going to get from you all today. You know, I had a lot of things in mind, but using AI to generate farm stand payment links for gourds was just not on the top of my list this afternoon, but I'm excited to see it. Okay, so I think, you know, one of the other things that I think, and I know this is running locally and we won't get into like sort of the local mobile debate right now, but I just think that the ability to do this stuff via text on the go is so powerful. Imagine somebody wants to buy the three pumpkins off the back of your truck and you want to be able to do this really easily, this sort of like ability to just type in a short command to get a really custom software solution backed by, you know, either a platform or not, I think is pretty

cool for people. I think there's like so many little details too that again, like while you're watching it, It's a little slow, but it just figures this out for you without any intervention. Like, oh, it has to decide which location I'm selling these from for the payment link. And it just, it just like saw that. It needed it. And like when it looked up what locations I've got here. And like just it just unblocks itself, which is really nice.

So you can just wait for this to go. And so, okay, so I actually have a real payment link. And this is just going to load up. It uses my items, has a price. Look at that. Yeah. And so we could just throw in a real payment link. real credit card right now and this would work and we would collect on that.

And so this is like, okay, so this is a bit of a tour of an existing MCP, like how you connect these agents to real products. But then I think what's like part of my job is really like, okay, what's the next step and how do we make the agent do the next step? And so for me, I think we could switch to from here is like maybe we go give it another MCP so that it can send these payment links in an email. And so I have a test email account here that's totally empty. And I think we can work with that and then just try to show you how you can literally

like extend the existing conversation. So I'm going to leave this open and we'll come back to it. I'm going to make a new window that we're going to use to make the like vibe coding half of this. So we'll just go vibe code and MVP together. I'm also going to open up a terminal window so that we can follow along. And I did do one thing in advance, which is I went in great. grabbed an API key here in an N file. And I set that up just in advance so we didn't have to show the API key on the screen.

And so we're going to use that to set up an MCP server with Mailman. And I want to just pause here because we've heard this a lot on this podcast, which is people are always saying, you know, you can do the output. I'm sure you could take this link and send the email. But anytime you're about to create an output, instead think about how can I create a tool? or a prompt or an input that could do this for me. So instead of saying, great, I'm going to email this link. You say, great, I'm going to build an MCP to send emails. Exactly.

Yeah. Like, how do you take that whole end-to-end workflow and solve every part? Okay, so let's talk it into vibe coding at MCP for us. So the way I like to do this usually is I'll start with the scaffolding in making sure that everything kind of connects and works. So I won't even say MCP at first. I'm going to just say, like, let's try to send an email. So I'll tell it, like, I have.

a staff-holded hyphen package here using UV. I'll tell it about that dot-end file I mentioned. And I'll tell it what's in there because I don't want it to actually read this. It's like, you know, I don't want the model to know about my API fee. And so then I'll say, can you make a Python script in this directory that sends a test you know? And this is like the vast majority of the work. Like it going from a blank slate to sending a test, email, that's the hard part. The NCP half is really easy. So yeah, I always start that way. Like, let's just prove that everything works. That you can do the thing and then we can figure out how to invoke the tool. Yeah. And this is actually another feature of goose that I'm kind of glad it accidentally triggered is that it tried to read the dot-end file because it was like, oh, what am I

working with? And we denied that actually because it's like a security concern. So there's a bunch of stuff that we try to catch. And so it can't see it, but it's going to have to just trust me that what I said is in there and then use it. And so, okay, let me follow along over here. So I'm going to just, we have a new file now, send test email that should be using this. And it just wrote that. And if I wanted to like follow along in BKale, I could do, I could, I could read it in an editor or something like that. But for the most part, I'm just going to trust it and hope it works. And so let's say, okay, can you,

and I'm going to give it that demos email. Got it. So now it's written the script and you're just going to double check. This thing works and that I'm going to get an email out of the script. And so it figured out what I meant by the dot end. So it's like, okay, I found a thing. And it's got this hard-coded recipient that it's going to go edit. And you can see like if I like just read gather.

They'd like just fix that for it now that I told it who it's going to send it to. And it's figuring out all this virtual and stuff. Okay. So just AI, just like us, figuring out its virtual environment. That is the theme of up to this. Yeah, every time. Okay. And then let me go over here and see if we got it.

And I'm sure it's going to end up in spam or something. Oh, there it is. And so we got a test email sent with Python. And it goes to spam because obviously this is all demo stuff. That's not surprising. And whenever I, you know, I get engineers asking me a lot, they're like, mailgun, it's well documented. I can write the thing.

But I think about the number of characters you wrote in Goose and just the number, like the typing that had to happen in the file. And I just think it has just saved you 15 minutes of time of like looking at the docs, copying it over, you know, checking that it works. And we did it in probably two minutes. Yeah, exactly. And like I just, I've, I've never built an email sending service. So it's like, I don't know this API at all. It didn't have to ask me for that. It's like, oh, it knew the endpoint name and everything. That's all trained in, which is just fantastic. Okay. So now that this is working, let's let's turn it into an MCP. And then this is an example of like a little bit of model knowledge going a long way. So this model isn't trained recently enough to know about the MCP SDKs. So this is an example. So, this model isn't trained recently enough to know about the MCP SDKs. So that's a thing where I'm going to actually go get it that context so that it can use it. And so I'll just go find like the, yeah, MCP Python SDK and we'll just grab it from the read me.

And so at this point, I could actually tell Goose the URL and like you could download this read me and it would work. But I'm rather just like copy and paste honestly because it's easier. I'm the same as you. I'm just like let's just drag the whole page and paste it in. Yeah, and it's like, okay, great. Like, it's already got these links. All I need is just this one example, and that should be enough. And so I'll say, you know, okay, be nice to your LLMs, of course.

I agree. Be nice to your LLMs. So I'm going to say now we're going to make an NCP server out of this. Here's reference code. And then I'll say, can you use that example to make an NCP? with a base and email tool. And we'll see how far it gets with no other context, actually. We might need to tweak stuff, but hopefully that's like a good start for it.

And while it's thinking, are you doing, does Gooselect models? Do you have a default? Do you have a favorite? We currently don't auto select models because just people bring so many different models to Goose. And so for us, we kind of switch context by context like Jackie. So like for writing code and using these tools, currently I'm using Claude Sonnet, which is great. But when I, you know, if I wanted to let go write a design doc, I might actually switch to like one of the reasoner models from opening I. Or I could even like try out like a local only model and run QN.

Okay, so let's see what it's doing. I'll check in again. And I'm just going to like take a look. Okay, so it's got a new file and it's combining like what we were doing before and and it's following that example pretty well, what it looks like. And that's actually good code. I'm going to interrupt it honestly because I know what it's going. It's going down like a slightly wrong path because it doesn't know how I'm going to integrate it.

So I'm just going to pop that. I like the judgment. Yeah. The code. It's like definitely the right idea. I'm not going to like read line by line because I kind of trust it at this point. But what I am going to do is I'll make my one line contribution to this project right now, which is I know I wanted to be running not as a remote URL.

I want it to just be a standard I.O. MCP if you work in it, that's the difference here. So I'm just going to make it standard I.O. And then I'm going to ask it like a weird question to help me set this up in Goose. So I'm going to say to help me do that, can you get the path in this virtual end? So really, I'm just trying to make sure that I have a portable command so that I can run in use. And so I'm going to ask it to figure out that virtual lens. So listen that. And that should just take like one or two tool calls for me.

And is this your just general flow for making MCPs all the time? Build me the function. Then here's the code to turn it to an MCP. I'm going to make my one line contribution. I'm going to ask this very specific question that I know works in my system. And then you're going to get it, get it going. Yes, exactly. And so once you have that like a core functionality working, then you give it this example and turn it into the MCP. Exactly that. Yeah. And so, okay, so it got me this, like this is the path of the virtual length Python, which is what I needed. And then I'll ask it like also, what's the abstract of the Tycon file? And that's going to be the two ingredients that I need to go make this work. And it's just, you know, running shell commands to do this. So I've already got it there. So let's do this and I'm going to go

into that like existing settings and I'm just going to immediately add a new custom extension that we just built. So I'm going to call this email. We're going to do that path to the Python file. And I didn't paste the other one. So let me just go to the terminal and grab this. And there and we're going to do we're just going to run that server. We just created. And that should be everything. Oops, I got an error. Let's see what that is. I'm just going to paste the output.

Oh, so there's actually a code error. Oh, well, there you go. And I'm going... It wasn't good code. So I'm going to tell it this... I'm going to tell it this error and see if they can just fix it. So I like that you actually say I got an error. Can you help me?

I just go straight to just paste the error in. Yeah, it's probably sparting. I like spend a little too much time, like, having a conversation, honestly, with the LOMs. And so, like, hopefully you can figure this out without me thinking too hard about what's in the file. We can, like, watch it work, of course. And honestly, I think the thing that it did here is this resource. So it's fixing that. but yeah.

Okay, so it figured it out at about the same time I did. Okay. This is going to be the new podcast. It's human developer, AI developer. Who figures it out? Who spots it first? The debugging flow is like something that I thought was going to be what kept me employed with the AI writing all these code.

But I'm like, oh, it's actually pretty good. It's pretty fast. Yeah. So we're going to try that again now that we've edited that file. I'm going to interrupt it because, again, I think it's doing some extra stuff we don't need. And I'm just going to toggle that. Where did it end up in the list? I'm going to toggle that on you.

Oops. Okay, new error. Sorry, I don't that. Normally this works in one shot, actually. Do you find as an engineer that you have shifted your debugging strategy to this, which is just I'd rather loop with the agent and just get it to figure it out than sit here and use my eyeballs and my fingers to solve this thing. I'm just curious how that's happened internally to yourself. I think I almost always let the LLM try once.

Like if it knows what to do to fix it, I'm like, great, that's faster than me figuring out. And then the thing that I think I am better at than it is is like just being tenacious. So if it's like if it's not figuring it out, it's like, okay, I will take over. And so this version is like, I'll take a look at like whatever idea it's coming up with. And if it like seems like it's in the right direction, I'll let it finish. We did it. Hooray. Great.

So now that we've got the like infars loaded. We actually, so what we just did is we like turned that thing we just built on in the existing chat. Okay. So it's, it's an ex. It actually should now see it. So I'm going to say like, great. We just enabled that. Can you see the new tools?

Try sending a test you know. And let's see if that works. So now we've at least like got it running. And yeah, it sees the tools, right? And it's trying this like get email status, which doesn't, we don't need that. But it's like, great. It sees that everything is ready to go. And then it's like sending me an NTP email test.

And it is success. And so it's working. And so we kind of built it, plugged it into Goose and then immediately tested it out. And let me go check my inevitable spam folder. Yep. There it is. You did it. Great.

So despite all the spam checks, we've got that working. And then now we can kind of finish that off by bringing it. it back to where we started, right? So we know that works. I'm going to go over to this existing chat. And these are two separate conversations, right? So like this this LLM knew all about everything we just did. I'm going to close that out. This one has no idea that that just happened. So I'm just going to go in and toggle that on and off real quick. And we're going to go back here. And I'm going to say, I just enabled an email tool. Can you send that payment link? to my demo email address.

So it doesn't have any of that, like, it doesn't have to, its memory doesn't have to worry about it. It doesn't know the functions or any of that stuff. So this is the real test. Yes. So it has no idea other than that this thing just got created. And it's doing the same thing that the other one did, which is like, okay, it's like checking and knows where it's coming from.

And yeah, hopefully we'll get it to send something. At this point, it's like writing more content, right? because it's like, I'm sure it's going to try to be cute about the, the, like, body. Are we going to get a pumpkin emoji in the, right. Yeah, that's what I'm open for. Okay. And we should have something in spam. Yep.

And so we've got produce store payment. We've got our payment link, order details, all of that stuff. And, you know, the same link should bring us back. So there are two levels of faces I make on this podcast. This is one of them. I go, ah, and then this is the other one where something works. So you've gotten the raised hands podcast reaction. So just to recap everything Goose did for us today and like still processing.

Jackie, we got your CSV. We took some like dirty produce data, got some insights on how to make it better. Brad, you took that and populated a store with your produce items. Then created a way for people to buy those items through a nice little friendly link. Then you vibe coded an MCP live, thrilling, as somebody who's avoided building an MCP for no less than six weeks. Now I feel like I have no no excuse other than to just go do it. And then you use that MCP to send somebody a pumpkin purchase email. That's right.

This is so great. Okay, this has been incredibly fun. I have to ask you both as part of our lightning round questions. Jackie, what's your favorite MCP tool in your stack right now? I hate to say it, but it's probably Google Drive. That is as a manager where I live. And I'm just so happy to be able to not have to copy and paste and click and save the a ton of time.

Okay, Brad, what's your favorite? Definitely Shell. I just, you know, I've lost all my memory of the arguments to various shell command. It's just 100% filling them in for me at this point. Okay. And then if you could make a pitch to people inside an organization that are hesitant about getting started with AI or even people that are outside of technology that just say, I don't know, how could this apply to me? Why would I use it? Seems really complicated.

What would you say to them? Find the thing that you don't like doing. automate that. Like don't, don't, if you love programming, don't try to automate your programming. That's like fun for you. You maybe figure out when, when you want to make that trade off. But it's like, oh, I hate writing the unit tests. That's the thing. I'm going to have the AI do. Okay. Get rid of toil. Spoken like a true, true engineer. Jackie, what about you? I think really treating it as a learning experience and realizing that this is the worst the models are ever going to be. And so even if it's not something that transforms your daily life right now,

even by keeping up with the tools and seeing where it fails. Seeing where it fails is like just as useful, seeing where it succeeds and still really training as an experiment in a way to learn versus, you know, a magic bullet that's going to magically solve everything. Okay. And then Brad, we saw how polite you are. We have to ask both of you our favorite question, which is when AI is not doing what you want and you're starting, maybe you get exasperated?

I don't know. What's your tactic? You seem very polite. So every, I saw every line started with, great, but you really screwed that up. So, Brad, what's your strategy here? So I thought this was kind, but I realized now it's even more brutal, is that like the moment it's not good. I just throw out the session.

I throw out the whole combo and I start over. And it's like, it's so much work to convince it to do the thing you want. It's so much easier to start over. And then I'm like, oh, I guess that thread with that model, it's just done forever now. So it's kind of sad. Can you imagine if we did this at work where somebody started to say something that was totally wrong? You just like shut the door in their face and said, come back and just start. Start totally, totally over.

Okay, Jackie, what about you? When it's not functioning right, definitely start a new session. But what I like to do is have it. I'll give it like really keywords that I want, if I want the style of something to change. So, you know, if something comes out like to MBA, like a lot of times, I'll say write it like you're a hacker. Another thing I like to do is how to summarize everything that it did the good parts and then take out the bad parts and start a new session based on those instructions. Great. I have learned so much myself. I'm going to go download Goose and build that MCP that I've been avoiding. Where can we find more information about this? Where can we find you two? And how can we be helpful?

Yeah, definitely check out block.com.github.io slash Goose. That's like the homepage, all the install instructions. and it should lead you also to our like GitHub repo where we love contributions. And if you want to chat with us about it, we run a Discord for Block Open Source. So get in there and like all of us are in there and we like debug live and all that kind of stuff. So a lot of activity for sure. Amazing. Well, thank you so much for sharing all this on the show. Thanks for having us.

This is fun. Thanks so much for watching. If you enjoyed this show, please like and subscribe here on YouTube or even better, leave us a comment with your thoughts. You can also find this podcast on Apple Podcasts, Spotify, or your favorite podcast app. Please consider leaving us a rating and review, which will help others find the show. You can see all our episodes and learn more about the show at how IAIIPod.com. See you next time.