What I’m thinking about: Watching my own browser pull comps and underwrite a live deal on its own…and why the leap that made that possible is showing us the future of software.
Same engine, a new way to deliver it
As an update on SLC Chat (our AI underwriting brain, trained on thousands of land deals), the first version lived on a Replit backend as a public web app. The engine was solid, but the friction killed it: you had to go find all the comps yourself, or source an existing diligence packet, and THEN feed it into the app to spur the pricing function.
As of 2026, with the growing agentic capabilities of tools like Claude Code or Perplexity Computer, nobody (including myself) has the patience for that. People want to throw a property into SLC Chat and let it run (e.g. finding all necessary comps), start to finish. As Hormozi often notes, out of any competitive advantage, speed usually wins out.
So we moved the web app functionality into a Cowork plugin that handles the whole agentic process, allowing for automated comp selection.
The underwriting logic, the pricing mechanics, the way it’s tuned to how I actually think about a deal… all of that was already built and battle-tested. None of it got thrown out. What changed was the delivery.
(Of course, SLC Chat also has built-in logic for how we approach different markets, our current thoughts on the macro, and how we utilize AI in our business, as additional examples. But the land pricing function was by far the biggest driver of interest, so we had to focus our delivery around that).
A couple of other advantages were that the token cost shifts to end user (the customer’s Claude subscription subsidizes the usage, much more sustainable and cost-effective compared to us having to mark up the web app price significantly to account for expensive Claude API tokens), and we are able to utilize Anthropic’s validated agentic backend that lives inside Cowork, instead of attempting to build that software inside a vibe-coding app like Replit (likely to be far less reliable, and a lot more expensive).
The workaround that protects what makes us, us
Perhaps the single biggest disadvantage of creating a Claude Cowork/Code delivery mechanism is that the source files for the software live on the end user’s machine, exposing valuable IP like
our proprietary pricing math (which utilizes hundreds of variables, and was refined over years with millions of dollars of capital deployed).
While we realized that ~99% of potential customers would not attempt (or understand how) to reverse engineer the source code for SLC Chat, the edge case gave us trepidation as a major tradeoff, with no route to put the genie back in the bottle. (There would be minimal to zero ‘paying access’ protection too, as even if we offered routine updates to the plugin for a recurring sub, if a customer stopped paying, they could run a previous version indefinitely).
The fix we discovered was to create a custom MCP connection (the layer that lets AI talk to another piece of software) that keeps the core pricing math OFF the user’s machine entirely.
The Cowork plugin runs on the user’s side, while the proprietary logic stays gated on ours, and because we control access, we control delivery (e.g. each user has their own API key assigned to them, giving us payment protection, since we can remove API keys, shutting down the user’s access to the core functionality of the software).
To me, this is the future of software. Boutique, white-glove tools installed directly on the end user’s workstation (individualized to a ‘T’), usage cost on the user, the IP protected for next to nothing. Personally, I haven’t seen anyone else build something like this yet, but I’d be stunned if we don’t see this as a growing trend.
Why it runs this reliably
Anyone could open Cowork right now, point it at a parcel, saying ‘comp this,’ and the results would almost assuredly be disappointing.
Comping land is one of the hardest things in this business to standardize (a big reason land has stayed a blue ocean this long, even in a down market), regardless of whether you are teaching another human or a machine.
The guidelines have to be airtight, and the QA process continuous.
What carries SLC Chat is EVERYTHING we’ve already built: the search-and-filter logic for MLS listings (accounting for granular listing history), the physical characteristic matching, the exact pricing mechanics. All of it validated across millions of dollars of deals and codified during the ~$100K, hundreds-of-hours Land Pricer AI build over the previous couple of years.
That data is the edge. As a reminder for how Sharran Srivatsaa puts it: amateurs automate for efficiency, professionals automate for accuracy.
Knowing when a comp actually belongs, when PPA stops being reliable on a small parcel… that took years, and it’s the part you can’t prompt your way into. Earned judgment and expertise will never cease to be valuable.
Watching the computer do the work
I recently tested the Cowork plugin on a live, non-disclosure Texas property (I wanted a ton of difficult variables thrown at the model, not an easy layup).
First, all I did was tell Cowork to, “Comp and price this land parcel, APN [123] in XYZ County, State,” and then I sat back and watched my browser drive itself.
Claude in Chrome (Claude’s agentic web browser, which was infamously janky and unreliable until roughly a month ago) jumped between LandID (identifying the subject parcel on its own, and using overlays to log characteristics) and Redfin on its own, correctly setting the search window for comps, cross-checking characteristics between the subject and comps, and throwing out the comps that didn’t fit. My computer, working a deal without ME touching it…exactly as intended.
As if this wasn’t impressive enough, mid-run, Claude built a faster version of the plugin, an extractor tool to rip through pricing history across every comp (The extractor is saved for a future release while we lock the exact pricing specs first… given the risk of patching technical debt into the software).
It’s certainly not perfect out of the gate, but it is breathtaking how Claude holds the entire multi-step SLC Chat plugin process (thousands of words of nuanced instructions), figures out the efficient path, and keeps moving.
The entire comping and pricing process takes ~20-25 minutes (likely to decrease as we refine the backend), and it needs zero attention from the user. That’s already a major win: token cost is limited on a Claude Max plan, and you could easily run ~5-10+ of these in parallel.
During the testing phase, I’m keeping a careful eye on the process, flagging what to sharpen, patching multiple fixes at once, reinstalling…and it comes back tighter EVERY single run.
It still feels surreal how reliable the tech finally is (again, this would NOT have been feasible even 5-6 weeks ago)…enough so that this product development is a refining process now, not a fight.
It’s like having a full team of engineers cleaning up behind me, endlessly available and for almost no cost. A year ago, getting here meant serious money and outside help (exactly what I did with Land Pricer, and it’s funny how sown seeds can sometimes be harvested in the most unexpected ways, later than ever imagined). Now I run the entire build process myself.
I’ll demo this live soon…watching the browser work itself, and seeing the tool rewrite itself mid-run, is the part that never lands in a screenshot.
This is the most fired up I’ve been about a tool in a long time. As a business owner who has consistently added complexity to our operations, often to our detriment, I’m laser-focused on simplifying our processes.
Ultimately, that should be the entire point of what we do day to day: build the systems that automate us out of the operations…leaving room purely for creativity and non-rushed decision-making.
If you’re imagining the possibilities of a tool like this in your business, drop me a note, and I’ll add you to our priority list for access (again, this is a white-glove install, built uniquely for YOUR business, so we’ll only be able to implement a handful of these plugins per week).
=====
Never forget that tools like the above are all built to serve our underlying deal engine: We write checks from $50K+, close 100% of the deals we commit to, and bring national underwriting built across thousands of reviewed transactions, with millions of dollars deployed.
