In this episode, the rebuild of SLC chat from a Replit-hosted web app into a client-side co-work plugin gets a full teardown, tested live on a non-disclosure Texas parcel. The breakthrough is routing proprietary pricing math through a custom MCP connection on a near-free Cloudflare account, which keeps the IP off the user’s machine while shifting all AI token costs onto the end customer. This sits on top of a business that has funded over $6.5 million in vacant land deals at 41% operating margins.
Key Takeaways:
- MCP Guards the IP Proprietary pricing math sits on a near-free Cloudflare MCP connection with a per-customer API key, so it never touches the user’s machine and cannot be jailbroken.
- Costs Shift to the End User Rebuilding as a client-side co-work plugin instead of a Replit-hosted app puts all AI token usage on the customer’s subsidized plan, erasing SLC’s per-use cost.
- The Moat Is Data, Not Models The real edge is years of validated comps and the exact pricing specs from the $100K Land Pricer build, not the AI that anyone can run.
- Live Test on a Hard Parcel A non-disclosure Texas property ran in 20-25 minutes across LandID, Regrid, and Redfin, with each error logged in Notion and patched per iteration.
- Agents Build Their Own Tools Co-work live-built a pricing-history extractor mid-run, now slated for the plugin itself once pricing specs are locked to avoid technical debt.
Listen to the full episode for the exact LandID-to-Redfin comping workflow and the MCP setup that could reshape how boutique software protects its IP.
(Podcast transcript below)
Welcome to Get Serious. We’re at Serious Land Capital. We have successfully funded over six and a half million dollars worth of vacant land deals with industry leading 41% operating margins. So today I wanted to give a very exciting update when it comes to our SLC chat, co-work plug-in progress. So, you know, you’ve heard me talk a number of times before on building out this
more recent web app been yeah, you know, over the past couple of years had worked on a separate business, land pricer, where the goal really was to build a effectively an automated pricing mechanism enhanced by AI for any vacant land in the US ultimately a as a separate business that was too ambitious to handle personally you can
Hear about me, you know, speak about that or write about that. You can check all the archives if if you’re interested. And now the Land Pricer project is in the hands of another operating team who have taken it in a enterprise focused direction for appraisal firms in the US here. I’m very excited about their progress too. It’s a whole nother discussion. but suffice to say they had an even better vision for what that company could be than
I I did with you know split attention between two different businesses. So the more recent version that I’ve gone over before is you know now that we have you know just some of these more high powered AI models out there, especially you know quad opus.
really once it started to hit you know the model 4.6 and and above here and able to handle a lot longer tasks. I got the idea, okay, we could take another stab at building out a more bare bones pricing mechanism for for land, but also to incorporate all of serious land capital’s underwriting expertise into a web.
Where you would effectively be able to fire off any potential property to get reviewed, or to kind of double check from our own perspective how to underwrite something, or just, you know, even our own thoughts on market conditions and so forth. and, you know, the primary use case, especially with some early testing and just what what we found to be in internally as well here too, is that.
the comping piece and trying to price out properties is really where a lot of folks would want to put most of the attention. and that’s you know less reliant on having super up-to-date market analysis and so forth, which you know we have. But as you well know, you know, macro conditions can change on a day-to-day basis or how we think about AI or how we’re even using AI. there are some general principles in place, but again, that
area is moving so quickly. whereas our core underwriting mechanics and how we price properties is far more stable and just you know all ultimately of keener interest to a lot of our audience. So the key issue is that when we built out the first version of SLC chat, you know, we built it with a Replit backend and put it for a publicly available web application.
but the issue is like, you know, this had come out post-co work, where there was already a lot of agenti processing and you know, people’s expectations for what software can do are just you know evolving by the week here. And it was you know, again, both from internal feedback, me using the app as well as external here. there was just, you know.
some unnecessary friction in terms of okay, yeah, I need to find all the comps first or kind of produce a due diligence report and then throw it into the web app for it to you know spit out the underwriting analysis. when in reality having HMT processing, people just want the simplest pathway forward. It’s like, hey, just throw the property in there and then have the the chatbot.
Do the work through agenda processing. So that was the direction we headed in. but we needed to build it with a co-work plug-in back end instead of via replit host. yeah, a lot of the reason for that is like if you’re trying to build an agenda processing with one of those vibe coding apps, I mean it’s gonna be less reliable and you know extremely expensive to do that as well.
and switching to just creating a plugin that can be installed on the user’s co-work you know client that that they have on on their own workstation first takes the cost of the usage for the app off of of us and you know that that that’s just the issue right now is you know ai tokens are just quite expensive and so
you know, for us to maintain even a margin that would you know just be breakeven was still gonna be pretty expensive both for us and the cost we would have to put pass on to
to the user as well. whereas again, if we can work with a cowork plugin, like all that usage in the tokens are going to be utilized by the end user or the end customer instead. plus they’ll be more subsidized with the current AI plans. You know, if you have a monthly, yearly subscription, what have you, you just get more bang for your buck in terms of token usage.
The trade-off here is that if we had to create a software that lived on the user system is that it would expose our IP because, you know, we couldn’t necessarily gate all of our information and mathematics on how we price properties and so forth, you know, through like a replit back end or, you know, our own hosted web app, it it it would live on the user’s system. Or so I thought.
we we actually found a workaround for that, which again, I think this is just absolute cutting edge, leaving edge capabilities that that we’ve really started exploring here. I haven’t seen anybody else do this. I’m sure there are there there are, but you know, the folks who are on like the true forefront of of AI, like it’s it’s kind of few and and far between. and obviously I don’t know all you know eight eight billion people on the planet here. So
you know, small circle you have noted. but I, you know, even amongst the the folks who are like the furthest along in the AI space, like I haven’t heard anybody do something like this specifically. And to me, it’s like setting up what I think the future of a lot of software applications could be. And
So really to jump back into the process here, you kind of got the back, the background, what the feedback was, you know, what direction we started to have to take this thing into. but very fortunately, like most of the work was already done. so like all of our underwriting knowledge, how to price properties, you know, how we view the market and so forth, how it it tuned, you know, how the model tuned itself to my voice specifically and all the nuance thrown in it, like all that testing was already done.
Now the next piece we had had to do is like, okay, can we build in that agentec processing? So you know, if I’m running a cowork plugin, can I enable Claude and Chrome to then start going on to land ID and Redfin and start pulling up the correct property and how to comp properly and so forth. so obviously, like anybody who’s run comping before, like that can introduce tremendous amounts of error and
Yeah, I would say most people are you know still struggling to do that very effectively. And you know, we we would include ourselves on that. Like, you know, pumping land oftentimes is a difficult process, a lot of art to it, and hard to fully standardize, which is a reason there there’s such a huge learning curve to this industry and why, you know, it’s continued to be a blue ocean for for for such a long, long period of time.
And again, we we had a a big leg up is because like we already had all these instructions on how to properly search and filter criteria on Redfin or Zillow or Land.com, what have you, how to properly find properties on land ID and assess their characteristics and so we’re like all of that was in our original land pricer you know guidelines that we built for our initial software here. So
you know, e even though it took a lot longer at a much higher cost than anticipated, to you know get to a point where it might become more valuable to the marketplace. Like all that work we had done, you know, the hundreds of hours spent and you know, 100K of personal expense building out the initial version of the of the land price or product.
Like all of that could be plug and play into this new software that we were building because like we already outlined all of the processes and how to search and filter and what the pricing mechanics are for the land anyway. Whereas, you know, like you you might hear me saying this, like, yeah, I can just build something on my own. Like you could, you know, everybody certainly could you know, just go instruct cowork, okay. Like I need to find comps for property.
I I think you will be pretty disappointed by you know what you might get out of the box here. Like the guidelines have to be super tight. I mean, again, it it it it co work or not, whether you’re trying to teach somebody on your team, another human on how to comp anyway, like anybody who’s done that before handle the training. especially the ones who really understand how to underwrite properties, which I would still say is a minority, like a like a single digit percentage of the land industry.
really, really under understands how to accurately value land. like if you teach somebody else, like it is not an easy process, right? Like there’s so much nuance involved to do with it and so much variability on how people might determine what’s an appropriate comp versus not, especially when you’re considering you have tons of different markets in the entire US geographically. So that has to be accounted for
as well here. whereas like we already really trusted our process. It’s been validated over millions of dollars of of deals all across the country. And we could, you know, effectively plug that in to train co work to handle that through its agent processes. And so when I initially handed this over for Cowork to to build out again like did these news models like the op Opus four point eight Max like it’s just unreal the context window that it can
you know, handle so many different steps, sequentially, not necessarily in parallel, but still very, very quickly. and huge, huge requests that I’m like multi-hundreds of words of, you know, multi-step nuanced processing that I want co work to be able to handle. Like it it is just able to ingest that and not lose the thread. sometimes, yeah, it’s not gonna be a hundred percent.
Perfect and then like it might miss something here or there, get tied up on a more minor thing. But for the most part, these new models are just unreal in terms of the accuracy and yeah, accuracy plus just the reliability of being able to stick with a process and figure out okay, what’s the most efficient and effective solution here in order to produce.
an engineer something that the user is trying to achieve. And so again, like I I have to reiterate, you know, I could just shout from the rooftops here is like we already had so much inherent data and instruction on what to feed in the cowork that made this a lot easier. If you’re trying to do this from scratch, it would be a lot more difficult. no matter how good these these models are. And then context is king when it comes to
getting qual quality results out of AI. so I was able to plug all that in and you know said okay yeah the goal here is to create a co-work plugin that’s gonna have to be installed on on the client side. And it was able to start building all of this you know the process okay I’m gonna build out the plugin here here’s the instructions on how Cloud and Chrome is going to operate and so forth.
And I did mention, okay, yeah, there there’s key IP risk here. It’s like we’ve you know, we we would love to have more protections when it comes to our proprietary math on how we account for you know literally hundreds of different of of of variables across any potential land property and you know how we account for days on market or estimate non-disclosure sales prices and so forth.
like all of that we would prefer to have guardrails on instead of just handing it to the user. And even if like we did really obfuscate it, like hit it deep into certain files, most people would not be bothered to go check in on that. Like it would still be like even myself diving into various files, like you just your brain starts to just get o overloaded and it’s really hard to parse out. Sure, you could use AI to help assist with that, but most people would not do that. E even so we would
you know, want to put some guardrails in place. And what Coworkers able to come up with is like, okay, just create a custom MCP connection. And you know, MCP connections, again, if you’re not familiar, just a way for AI models to connect to a different software effectively and that, you know, have the softwares talk back and forth with each other, sometimes with editing capabilities, on on one end, sometimes just reading.
what what’s on the the back end of of a different software. but in this case it was like, yeah, you can just create a separate MCP connection and all you have to do is set up this separate Cloudflare Flare account, which you know is effectively free for what you’re trying to do. And you stick some of the project knowledge into that MCP connection and you create a specified API key for each customer so that
their cowork plugin doesn’t actually have your proprietary info hosted on their computer. And and you could do this for as much as you want. You could host everything on this MCP connector. For a lot of it, like it’s not it’s not really necessary because it’s not necessarily proprietary. And it’s easier it to just you know install something and and a little bit faster on on the user’s and and system here but
Theoretically you could guard as much as you’d like, but like our core math, the things that we wanted to guard the most, yeah, we could throw it on the MCP connection. And so now the user would just have to call that whenever they’re trying to use the the plugin. and we guard the API access. So, you know, if certain payments didn’t come through and so forth, like we could revoke that. like it gives us more control over.
the product that we’re delivering in the in the the first place here and it keeps our proprietary info completely guarded because no matter how much the user might ask, hey, like I want all the the math info, like try to jailbreak it, try try to red team the the system, it’s not able to pull it because it it doesn’t actually have access to that that info on the host system. So like it’s completely controlled from from our setup.
internally here. So really, really cool setup here, which is like what got me more excited than almost anything. Because like again, to me, this is like the future of software here, where all of a sudden you can put the cost of usage on the end user here and you can create way more individualized setups. Like in this cohort plugin that we were building out, it can account for, hey, does this
land investor work in the Pacific Northwest, or do they only work in a certain state? Do they only do hunting properties and so forth? Like you can orient everything to what type of properties and areas that the user works in, you know, what type of discount they throw on a properties, all that can be individualized to that specific business. but the purveyor of the software, like us for instance, here,
still gets to keep the proprietary data safe at at the cost of like nickels or or virtually free. So all of a sudden now we have a more protected product from from our end that we can more reliably distribute without having to worry about you know copy cats or you know and and and so forth here kind of like devaluing our our offer.
over time because regarding all of that key key information. so very, very cool setup there. I know that got a little bit tech heavy, but hopefully you can kind of understand the process and and like to me how mind blowing this is for potential future software installs for a whole bunch of companies because all like you now you can just create boutique products as many as you would like
you know, h handle the full white glove install here, but still control the the ultimate delivery and potentially shut off, you know, depending on payments or, you know, abuse of the system, whatever. You know, hopefully not the case. You can also throw in various paperwork and so forth design and you know, just make sure you you have alignment here with who you’re you’re conducting business with. but that was an absolute massive breakthrough.
So that is a core piece there. you know, the next piece we we could spend so much time on this, but you know, it was just addition additional QA, you know, kind of quality assessment of the actual product here. And plus just the cowork plug in build by itself. it’s not incredibly user friendly. Like it’s still pretty like if for installing a full on software, again, like I it it’s stunning that.
You know, even a year ago and so forth, I was spending tens of thousands of dollars trying to build up my pricer here. And now I can just do it on my own with this these agentic systems and be able to create these white glove installs here. You know, still a learning curve on it. That’s why I’m offering it more of a white glove install and anybody who is trying to build their own cohort plugins. Like it it is not a overly simple.
process. Like it it is and it it it is and it isn’t, but you do have to use things like terminal setups and understand how to, you know, kind of talk computer language a little bit. Yeah, you can have Claude or ChetGT whatever guide you through it, but it can be a little scary technically here. So like there there certainly is an on-ramp if you wanted to do something like this your yourself here and I’m like not certain
How quickly that’s going to become super easy for everybody to do. Like you saw me write about last week as well, where it’s like, I actually think the learning curve has gotten steeper with AI over time, which is just counterintuitive, even as the models have gotten much better. But this is just another another example of that. So, you know, got through that whole process. And even if you have to install and reinstall and so forth, like it’s a little bit more involved there.
but then when it came to the process of actually reviewing land through this cowork plugin, like it was just absolutely remarkable. I mean, f first off, there there there were definitely some key issues here. And like I had just had my notion subtask list. And what I did, you know, is I threw the cowork plugin on and gave it a live deal. I’m like, okay, go go try to review this. One that I had already seen, like I’m like, I want to use this for my own business. Like this, like I I’ve got
You know, a four-week old boy at home, three and a half year old two, like I am just looking for every simple and easier solution. Like that that that’s generally where you’re hitting on your best ideas as as an entrepreneur is like, hey, the time sink and the you know laziness is just hitting an all all all-time peak here and just tiredness. Give me the simplest solution to run my business as effectively as possible. And this is a critical element here. Like if I can just throw in all the deals that come our way.
process them through the cowork plug-in, and then I can kind of assess on the back end here. It it’s just incredible how much how much time that’ll save. So I want to test it on on live processes. And this was a non-disclosure state at Texas property. So I want to test it like on something that was a bit a little bit more difficult with all of our pricing mechanics built in and so forth. And so I had my notion task list set up here and
after I threw it into the cowork side of things, I was looking at how Plot and Chrome was was handling it and just like, you know, filtering down the process. And I was just making notes. Okay. Yeah, there’s an error here. It could have you know, first it should have tried like if I just handed it an APN, like I not even an address, like an APN here. Don’t try to go to the county records here. You just utilize a logged in window of land ID or maybe the user has read grid. Just go find it in there and then you can
kind of reverse engineer what the nearest address is. So you can plug it into Redfin. By the way, do default Redfin versus Zillow. I prefer that instead. you know, the window aperture of where you’re trying to find land in Redfin, like it was too wide. So instead go back to land ID and try to you know use the measurement tool so that you can get a you know few mile radius instead to to make sure that you’re in the
correct vicinity for the underlying subject property and then only expand as you need to. And by the way, like here’s all the various filter, the the the filters that we use and the acreages and which you know, do we go to sold comms or pendings or actives first and so forth. like I’m naming a few examples here, but that that that’s just giving you an idea of like I was watching it go and like okay there’s a mistake, there’s a mistake there. I’m gonna note down that these issues. And so when it gives me that final results, I can plug it back into cowork and like okay, fix the software
patch the these mistakes so that it it it smooths out the process going forward and we get even more accurate pieces here like yeah didn’t read the pricing history here or for example like cowork as it was reviewing a whole bunch of comps initially it was going much slower and then it live built an extractor tool that could more quickly process the pricing history for all the comps that it was finding. So I was like, ooh that’s cool. I I don’t want
Co-work to have to burn tokens each and every time it goes through this process. Can we just reverse engineer this extractor tool to build that into the plugin itself? so now that’s part of the task list here, though we need to fully confirm the pricing specifications before that extractor tool is fully finalized, because otherwise you’re just
you know, kind of i improperly patching and just incurring more technical debt here. so like you know, paying attention live to how things were going and then just you know correcting along the way and then just okay I need this fix, this fix, this fix here, and just have cowork spin it back up. Like, okay, just patch, patch, patch, patch, patch, reinstall the plugin, and let’s run it again and let’s see if it if it goes even faster. And these initial runs were like, you know, 20 to 25 minutes.
But the timing really doesn’t matter here. Again, you could just and especially on these max plans here, like I thought, it’s gonna be pretty token intensive. My usage really is not that tough. And I’m using like the Opus 4.8 Max. yes, like Claude right now, they have some heavy subsidies, double the usage of trying to get more users on to their plans right now. I I I was accounting for that. Even if you know our usage gets cut in half here, like it’s still so minimal.
compared to the overall, like I I could run probably five of these at once, or it if not more. and and just you know parallel process a whole bunch of properties and then get that and report to review more more thoroughly here. So just such a cool setup here and like it it is just mind blowing, like see.
you know, my computer just and and browser just acting on its own and going from land ID and back to Redfin and like cross-comparing and making sure all the characteristics are matching and i excluding comps that that are not relevant here. Is it a hundred percent perfect that? No, definitely not. but like each iteration that I’m getting through is just knocking out so many bugs reliably. when I’m
you know, reviewing e each iteration. It’s just getting smoother and smoother at a time. It’s like just having a full team of engineers, you know, eight agents working on my behalf to clean up all these bugs and then just ready to push it out. again. and I’ll reiterate as well. Like anybody could do this. but do I think you could build something as reliably
without having those years worth of comps and history of reviewing properties all across the country and having built a previous software with all of the specs on how to do that exactly, volume mathematics on how to provide the exact same range, or not not the exact same, but the the precise range of pricing that each comp should be considered at when you compare it back to the subject property. That is where the magic is. Like I said, you know, anybody can do anything.
outside the box or you know out of the box with just you know your general ai models, but the data, the data on how to properly consider what that final outcome is and to to use proper judgment on what is right versus what is wrong because AI is going to be wrong a ton of times.
that that just goes to show of like understanding, hey, cowork is not going in the right direction here. We should, we should refine this piece here. Yeah, this pricing is not gonna be quite as accurate as we would like. Or, you know, if we’re doing acreage this small, we can’t really use PPA twice you know price per acre.
a as reliably here. So understanding how to make those tweaks and then build it back into reliable software as well too that you can deliver to others. that is where the real magic here is. And yeah, like I I can’t wait to try to release this more. Like that and now I’m really centering in on this like that this is going to be so major for our business. and each iteration I’m going through is just getting better. like you know
al almost a log better every single time here. it’s just absolutely remarkable. I I I cannot overstate how remarkable this technology is when it comes to being able to automate work that that again just used to take hours and days and you know months of slog to to get through.
but we are building the systems to you know automate ourselves out of major parts of the business, which is really all we’re trying to do on a day-by-day basis. Hopefully make some money along the way. so if you’re interested in this, you’ll you’ll be hearing more about it in both writings and your other webinars and demos and so forth that will be coming out here. my goal here, my
Initial deadline for the first version of this was the end of June. We are coming up here. So very excited by by the progress here. Hopefully this gives you an idea on what we were doing. and if you’re interested in this SLC chat co-work plugin, feel free to reach out. You got my email, Chris at serious land.capital, or reach out on the website if you have any deals you would like us to review. Maybe we’ll use our co-work plugin. and subscribe and share, everybody. Looking forward to next time. Take care.
