CityMorph

Inspiration

House hunting is stressful for anyone. If you use a wheelchair or a walker, it's much harder, because listings leave out exactly what you need to know. Here's what that search looks like today:

You filter for "accessible" and get almost nothing. In CoStar's own Houston data, only 72 of 1,619 listings carry that tag. The seller's agent sets it by ticking a box, so even those homes might have steps at the door.
The photos don't show what matters. You can see the kitchen counters, but not how many steps lead to the front door or whether the bathroom has room to turn around.
You find out in person. You drive to a showing and discover three steps at the entrance, or the only bathroom upstairs.
Nobody tells you which homes are close. A home with one step and a narrow bathroom door might need a few thousand dollars of work. You'd never know, or know that a state program might pay for it.

This isn't a niche problem. About 1 in 8 American adults has serious trouble walking or climbing stairs (CDC). Yet HUD found that only 0.15% of US homes are wheelchair accessible, about 1 in 600, and fewer than 4% have a step-free entrance, everything on one floor, and halls and doors wide enough for a walker or wheelchair (Harvard Joint Center for Housing Studies).

People in this situation move for the same reasons everyone does: a new job, rising rent, a parent moving in with family, or a current home that just can't be adapted, like one where every bedroom is up a flight of stairs. When your own home can be adapted, that's usually the best move. When you're moving anyway, you need a home that already works, or one that's close.

That's the part that got us excited. The same HUD study found that about 1 in 3 American homes is "potentially modifiable." The home a family needs often exists already. It just needs a ramp, a wider door, or a different bathroom, and no listing points them to it. We built CityMorph to find those homes and show exactly what it would take.

What it does

You start by telling CityMorph what you need, in your own words. Something like: "We're buying a home in Harris County for up to $350,000. My mom is moving in with us. She's 74, uses a walker, and can't manage steps."

CityMorph then checks every listing against that home's own 3D tour and floor plan. It looks at what actually decides whether a home works: whether everything is on one floor, whether there's a bedroom and a bathroom on the floor you walk in on, how big the bathroom is, how high the front door sits, how wide the doors are, and whether the listing's "accessible" tag matches what the tour shows. When a home fails on something it can measure, like a bedroom upstairs, CityMorph rules it out and says why. When it can't rule a home out, it lists the changes the home needs and the questions nobody has answered yet. On 40 real Houston listings, for a walker user with a $400,000 budget, it ruled out 16 homes and kept 23, each with three questions still open.

Next, it shows you the changes inside the home. The floor plan turns into a 3D model you can spin around, and you watch it change from before to after. A wall comes down to make room for a 5-foot turning circle. A ramp folds along the side of the house, with handrails on both sides. Handrails go up along the hallway, the tub becomes a curbless shower, and the kitchen gets a sink you can roll a wheelchair under. You can also see the ramp and the turning circle inside the real Matterport tour. Every change shows its size, a price range, and a link to where that price came from.

Then it finds who might help pay. It checks the household against the actual rules of programs like Texas's Amy Young grant (up to $22,500), the Texas Medicaid waivers, the VA's housing grants, FHA 203(k) loans, and the IRS medical-expense rules. For each one, it tells you what the program pays for, when it applies, and what it'll ask of you. CLASS, for example, needs three contractor bids for any change over $1,000, and the VA has to approve the plans before work starts. CityMorph always says a family "may qualify," because only the program can decide.

Some things no listing or scan can answer, like how many steps are at the door or what kind of shower the bathroom has. CityMorph writes those questions for the seller, and you approve each one before it goes out. When the seller replies, the answer replaces the estimate and the checks run again.

Finally, CityMorph puts it all on one page you can print or save as a PDF. It lists the changes and what they might cost, the programs and their steps, what you might pay after help, the things to check in person, and every source. Your case manager, the grant program, and your contractor can work from that page. It's a starting point for that conversation, not an assessment or legal advice.

Why you can trust the numbers

AI helps CityMorph understand what you need and suggest changes, but it never decides anything. Plain, tested code does every check and works out every size, price, and grant result. We wrote 830 tests for that code.

Every number on screen says where it came from: measured, estimated, asked, answered, or demo. Some things the data just can't know, like the true height of the front door or the shower type. CityMorph never pretends those pass. It labels them as estimates and asks.

Every program rule and every price links to its source, with the date we checked it. While re-checking our own sources tonight, we found a price credited to the wrong website and another with the wrong year, and we fixed both. The ramp sizes follow the 2010 ADA Standards: no steeper than 1 inch of rise per foot, a flat landing at least every 30 inches of rise, and handrails on both sides once the rise passes 6 inches.

How we built it

We built CityMorph with React, TypeScript, and Vite. All the decisions live in a TypeScript engine we wrote test first. A small server inside Vite runs the AI and reads the private data, so the API key and CoStar's listings never reach the browser.

The renovation check is the heart of it. A renovation is just a list of edits, like "remove this wall," "add a turning circle here," "widen that door," or "put a handrail on this wall." Our code checks each edit against the real floor plan. Is that wall inside the house? Does a 5-foot circle fit once it's gone? Is that door on the outside? If an edit doesn't work, the code rejects it with a one-line reason. If it does, the code updates the plan, prices the change, and works out where it goes in 3D.

We built the AI agent with LangChain's deepagents library. The main agent (Claude Opus) writes its own to-do list, searches and checks the homes, and hands the job of finding a better deal to a smaller sub-agent (Claude Sonnet). Then it plans the renovation for the person in front of it, walking through the home the way they would: in the front door, down the hall, into the bedroom, bathroom, and kitchen. It has to give a reason for every change. Last, it drafts the questions for the seller, matches the grants, and writes the family's plan into its own file system.

The agent can only use our 12 tested tools. Code writes every number in the plan, and if the agent's summary mentions a number that no tool gave it, we strip it out. Each run costs at most $1, and you can watch its steps appear live.

The deal finder works like a friend who knows the market. It loosens one thing at a time, such as one more step at the door for a walker user or a slightly bigger budget, and runs all the checks again. It only speaks up when the new option is truly better: the home works as is, it's the only option, or it costs less overall even with the extra money.

Where the data comes from

CoStar gave us Homes.com listings with prices, bedrooms, foundations, the listing's accessibility tags, and a link to each home's Matterport tour. We used 40 Houston homes. CoStar didn't give us permission to publish this data, so it stays in a private folder on each of our laptops and never goes into the code or the public build. Without that folder, the app runs on five sample homes that say they're samples.

Matterport's API gave us the inside of each home: the rooms, their sizes, the doorways, and a full floor plan with walls and doors. All 40 homes came back, with about 17 rooms and 77 walls each. We also used Matterport's SDK to place the ramp inside the real tour and to measure a door with two clicks.

FEMA's flood maps tell us each home's flood zone, which helps us estimate how high the front door sits. OpenStreetMap shows the streets, the road out during a flood, and the nearest ER and pharmacy. Houston's sidewalk survey and METRO's transit data tell us about the sidewalks around the home and whether paratransit serves it.

The grant rules come straight from the programs: Texas's housing agency (TDHCA), the state health agency's handbooks (HHSC), federal VA regulations, HUD, the IRS, and the 2010 ADA Standards. The prices come from national cost guides like Angi, HomeAdvisor, Fixr, and Homewyse. They're planning ranges, not quotes, and we say so.

Challenges we ran into

The hardest part was that our data doesn't know the most important things. No public source records how high a front door sits, and Houston's elevation certificates aren't searchable. A 3D scan measures a doorway's opening, not the space a wheelchair actually gets once the door is in the way. We decided "we don't know yet" had to be a real answer. CityMorph estimates, says it's an estimate, and asks, instead of letting a guess pass.

Keeping the AI honest took several tries. Our first version made up IDs for walls and doors, and later it suggested the same changes for every house. We fixed that by giving it a tool that reads the real floor plan, making it explain every change, rejecting changes that don't fit the plan, and checking every number in its summary.

We also had to use CoStar's data without ever publishing it, and we checked our prices by hand because most cost guides block automated reading.

Accomplishments that we're proud of

On real CoStar homes, the agent designed different changes for different people in the same house. For a walker user, it added handrails along the hallway, a taller toilet, and a non-slip floor. For a wheelchair user, it added a sink you can roll under, a lower counter, and a wider kitchen door, because it read in the scan that the door was 31.7 inches, just under the 32 a wheelchair needs. On a second house, it worked from that home's own layout and came up with a different plan.

We're also proud of what it refused to do. When nobody knew how high the front door was, our code wouldn't let it add a ramp of a made-up length, and the agent planned around that instead of guessing. Each run cost about 30 cents.

And every requirement and every price in the family's plan is a link you can click.

What we learned

We learned that an honest "we don't know yet, here's what to ask" helps a family more than a confident answer that turns out wrong. We also learned that an AI agent gets better when its tools are strict. Give it real IDs, make it explain itself, and let the code say no.

What's next for CityMorph

Next, we want to help people adapt the home they already live in, not just the one they're buying, by running the same checks and plan on their current home's tour. We also want CityMorph to look at the photos in each tour, so it can spot the shower type and the steps at the door on its own, while still asking the family to confirm. We'd like local contractor prices instead of national averages. Renters deserve help too, so we want to write the letter that asks a landlord for changes under the Fair Housing Act. For CoStar, this could become an accessibility layer on the tours Homes.com already hosts, so the one in three homes that could work stops being invisible.

Built with

TypeScript, React, Vite, Vitest, Node.js, Claude (Anthropic API), LangChain deepagents, LangGraph, Matterport Model API, Matterport SDK, CoStar hackathon kit (Homes.com), FEMA NFHL, OpenStreetMap, OSRM, MapLibre.

Sources for the statistics

About 1 in 8 adults with serious trouble walking or climbing stairs: CDC (Disability and Health Data System, 12.2%; earlier CDC release).
About one-third of homes potentially modifiable, fewer than 5% livable with moderate mobility difficulty, 0.15% wheelchair accessible: HUD, 2011 American Housing Survey analysis. We found the 0.15% figure only in others' copies of HUD's table, so confirm it in the HUD report before saying it on stage.
Fewer than 4% of homes with a step-free entrance, one-floor living, and wide halls and doors: Harvard Joint Center for Housing Studies, 2022.
72 of 1,619 Houston listings tagged accessible: our own count in the CoStar hackathon kit.