Inspiration
Scout for your new home or business. Listing sites answer "how big and how much." People actually decide on other things: is it near the school, will customers walk in, is the rent fair? CoStar has the data to answer those questions, spread across many products. We wanted one search box where you describe what you need and an agent does the legwork.
What it does
Scout finds homes for rent and commercial space for lease in Houston and Los Angeles from a plain-language description.
- Describe it your way: "2-bed with parking within 3 miles of an elementary school" or "coffee shop space downtown, under $4,000 a month." Scout reads your needs, asks when something is unclear, and never drops a requirement you stated.
- Get a shortlist: the agent picks its best three, each with short summaries built from real data: rent, size, distance to your places, fit-out, and nearby apartments and offices.
- Compare and explore: side by side with Vexcel aerials, Street View and Matterport 3D tours, plus a 3D furniture studio inside the scanned space.
- Know who to call: owner, management and leasing contacts on every property. 789 apartment communities, 1,700 floor plans, and 3,949 commercial spaces.
How we built it
- Agent: LangChain in Node. A typed tool call extracts your requirements first, code runs a validated search and checks the Matterport scan data, and the model ranks the candidates. Code validates every pick and writes the facts, so the model never invents them. OpenAI GPT-5.5 is primary, with automatic failover to OpenRouter, and fixed rules take over if ranking fails.
- Data and maps: Python pipelines over the CoStar kit, radius search on real coordinates, and landmarks found with the open-source Photon geocoder (OpenStreetMap).
- Visuals: the Matterport Model API and SDKs, the Vexcel ortho API, and the Google Maps Embed API.
- Prototype: a Python version built on LangChain DeepAgents. An orchestrator picked a top 10 and ran three deep-dive sub-agents in parallel, reporting customers, competitors and fair rent. Its business data preparation and profiles now power Scout's business mode.
Challenges we ran into
- Keeping the AI honest: we moved every fact into code and stopped the model from filling empty slots or relaxing your requirements.
- Reading intent: "studio" can mean an apartment or a yoga studio, and "near apartments" describes a location, not a home search.
- Unreliable services: overloaded OpenStreetMap servers and running out of model credit mid-test forced us to add retries and an honest fallback.
Accomplishments that we're proud of
- One sentence in, a validated shortlist out, for homes and businesses alike.
- All the sponsor data working together: CoStar, Matterport and Vexcel.
- Real 3D furniture staging inside real scanned spaces.
- An agent that won't make things up, and keeps working when the model doesn't.
What we learned
- Agents are only as trustworthy as their rails: structured tool calls and checks in code beat prompt wording.
- Deep agents shine when work splits naturally into parallel sub-tasks.
- "Unknown" builds more trust than a confident guess.
- Build for failure, and be willing to cut features users don't need.
What's next for Scout
- The full CoStar inventory, with live availability.
- "Will my layout fit?" for every scanned space: a 2D floor plan from the scan's room outlines with your zones and furniture laid out, and the same layout placed in the 3D tour.
- Clearance checks in the 3D furniture studio, so staged furniture is checked against walls and doorways instead of just shown.
- Drive time instead of straight-line miles.
- Foot traffic, demographics and how long similar businesses survived in an area.
- Deep-dive reports per pick from our DeepAgents prototype, inside the main app.
- Listing alerts and drafted messages to the right contact.
Built With
- claude
- codex
- costar
- javascript
- langchain
- matterport
- vexcel
