Private preview
Enter access PIN.
By invitation only.
An AI-native real estate website platform
The first real estate website that writes itself, ranks itself, and does the work — composed from a 38-slot library by Claude Opus 4.7, with compliance baked in from national fair housing down to brokerage policy.
The thesis
Every existing agent website platform — Placester, Luxury Presence, BoldTrail, Real Geeks, Sierra Interactive, Lofty, Savvy Studio — sells the same product in different paint: a pretty template, an IDX feed, a lead-capture form. The site sits there. Agents bolt a CRM on top to compensate for what the site doesn't do.
That model worked when "having a website" was the bar. The bar has moved. Buyers and sellers search through ChatGPT. Listings get summarized by Claude. Local relevance now requires real local depth, written fast, indexed correctly.
Stack is built for the agents who win that next decade — bespoke, hyperlocal, personal-brand-first, niche-deep. The platform recedes; the agent's work centers.
What it does
Neighborhood pages, blog posts, market reports, press releases, buyer and seller guides — every page composed from the slot library by Claude Opus 4.7. Compliance-passed before it ships. The agent prompts; the page exists.
Sitemap, robots, llms.txt, JSON-LD, IndexNow ping, GSC and Bing auto-submission — every page indexed correctly the moment it ships. Ranks on Google, gets cited by ChatGPT, Claude, and Perplexity.
Home value, market pulse, IDX search with any filter the MLS exposes, saved searches, CMA prompts — every interaction is a lead-capture conversation the visitor has with the site, not a form they avoid.
Where Stack fits
The competition splits cleanly. Real estate platforms own MLS data and compliance but ship templates without AI. Horizontal AI builders ship the AI but can't survive a fair-housing audit. Stack sits in the intersection that's currently empty.
| Cloud CMA | Placester | Sierra Interactive | Lovable | Stack | |
|---|---|---|---|---|---|
| AI-native page generation | · | · | · | ✓ | ✓ |
| Real estate compliance built in | partial | partial | partial | · | ✓ |
| Real MLS data integration | ✓ | ✓ | ✓ | · | ✓ |
| Custom per-site, not templated | · | · | · | ✓ | ✓ |
| Productivity tools in same surface | CMA only | · | · | · | ✓ |
| Pricing | $32–50/mo | $99–499/mo | $300+/mo | $25–100/mo credits | Credit-pack + brokerage tier |
Compliance "partial" = manual disclaimers, no automatic per-MLS cascade or fair-housing pipeline.
Why now
The full stack to build this didn't exist eighteen months ago. The combination that makes Stack possible — schema-strict AI, modern MLS APIs, edge multi-tenancy, prompt caching — only just became production-grade.
Public beta Nov 2025. Schema compliance is mathematical, not statistical. The AI cannot return an invalid page — the type system is the contract. This is what makes the locked-slot architecture viable.
Modern API + native MCP server turns 800+ MLS data sources into a 24-hour onboarding instead of a 40-90 day RETS slog. Verified live: 85 MLSs across 32 US states are addressable on day one.
Custom-domain SSL provisioning is one API call. Edge-first means margins survive at indie scale. Tenant sites land at their own domain in minutes, not weeks.
What would have been a 100-person engineering team is now a one-agent operation. Slot vocabulary, compliance manifest, system prompt — all cached. Per-page generation cost stays under a dime.
How a page gets made
The pipeline is short by design. Each stage has a contract — schema in, schema out — so failures localize and nothing escapes the compliance layer.
Page generation pipeline
Agent types: "Build me an Alger Heights SFH $200-300k page with a stat bar at top."
Claude Opus 4.7 with strict structured outputs + Repliers MCP tools. Returns JSON conforming to PageSchema.
Stage 1: deterministic blocklist (per-state). Stage 2: Haiku judge for steering, superlatives, claims. Verdict: pass · caution · block.
Pure function: schema → React → TSX. Auto-injects 9 locked compliance slots the AI never touched.
Cloudflare cache invalidation. Audit log appended with prompt hash, model, manifest version, verdict.
Sitemap update, IndexNow ping (Bing/Yandex), GSC submission. Live with full SEO + AEO meta.
Schema-as-memory: the page schema itself is passed every turn — no conversation history, no drift. Multi-turn edits ("make the hero bigger") use the same pipeline against the existing schema. Shipped as Plan 5A on 2026-05-25.
The vocabulary
No templates. The AI composes pages from a typed slot library — nine locked compliance slots the AI cannot touch, twenty-nine primitives it composes freely. Each slot ships with 3–5 visual variants. The combinatorial space across a typical 8-15 page agent site exceeds the variation across the entire US agent-site market today.
No schema representation. Injected by the renderer based on tenant state, MLS, and brokerage. The AI cannot remove what it never gets to reference.
The compliance moat
When HUD publishes new guidance, when California passes a new disclosure law, when an MLS rewrites a disclaimer, the manifest version bumps and every tenant site receives the new content on its next render. Zero customer action. Every render is audit-logged with the active manifest version.
Every AI tool in real estate has a compliance problem. Stack is the first one with a compliance thesis. Architectural invariants beat content moderation. We've already structured the platform so we can offer institutional partners a Master Compliance Agreement that covers every agent under their umbrella with audit logs, manifest versioning, and per-render compliance verdicts. That's not a feature we're adding. It's the foundation we shipped on.
The data layer
Data-fluent codegen
Most platforms bolt MLS data onto a page after generation. Stack's codegen calls Repliers MCP directly during generation. The agent types build me an Alger Heights SFH $200-300k page with stat bar and twenty seconds later the page exists with real comps, real median price, real days-on-market, real school data, and a filterable IDX grid configured for exactly that slice of inventory.
That's not a template that gets data injected. It's the agent's local knowledge plus live market data plus AI composition, all in one motion.
Hyperlocal by design
The agent's local knowledge is the missing ingredient in every generic AI site builder. Stack captures it at onboarding — voice profile, brand kit, market expertise, niche specializations — and feeds it back into every generation. The neighborhood page about Alger Heights doesn't just list comps. It reflects what the agent knows about Alger Heights, in their voice, ranked for the queries they actually want to win.
Each page is programmatically indexed at publish — sitemap update, IndexNow ping, GSC and Bing submission. Auto-healing SEO detects broken links, missing meta, invalid JSON-LD, and crawl errors, and fixes them. Every page, every query, every citation in ChatGPT and Claude tracked. The agent never thinks about indexing again.
CTAs in the right places
Lead capture is treated as a render-time decision, not a designer's afterthought. The compliance pipeline knows when an AI tool captures visitor data and auto-injects TCPAConsent. The schema knows when a page is for a high-intent search (luxury homes, Spring Lake) versus top-of-funnel browse (Grand Rapids neighborhoods). CTAs surface accordingly — primary on listing detail and market-report pages, soft on neighborhood guides, sticky-bottom on long-form posts.
The agent doesn't design conversion. The platform does. The agent designs their brand.
What's coming
Stack's v0 is the AI-native website builder with IDX and a starter suite of tools. Each subsequent release adds capability that compounds on top — the agent gets more product, the platform deepens its moat, every prior shipped feature gets sharper as the surrounding context grows.
Multi-tenant signup, schema-based codegen, 38-slot library, compliance pipeline, Repliers IDX with dynamic filters, programmatic AEO content engine, auto-indexing, brokerage BOR dashboard, in-product AI concierge, credit-pack pricing. Free tier for solo agents; $499/mo brokerage tier.
CMA generator, inspection analyzer, appraisal challenger, and listing-prep tools land inside the agent admin. Each is a standalone wedge product on its own; together they form the productivity layer the agent never has to leave the platform for.
Real estate has dozens of niche vendors — TransactionDesk, Brokermint, ShowingTime, FUB, kvCORE. Agents upload an OpenAPI spec, enter their key, and our MCP gateway auto-generates a sandboxed, security-validated server. Agents bring their own integrations. Lovable doesn't have this. No one in real estate does.
An always-on AI consultant that reads the agent's Follow Up Boss data, audits the pipeline, drafts the follow-ups, and surfaces today's priority list. Already working live as a standalone product against any FUB account. Embedded into Stack, it becomes the back-office the agent never opens a CRM tab for again.
A formal partnership program with NAR, individual MLSs, and large brokerages. They get audit-grade compliance reporting across every agent under their umbrella; Stack gets distribution, credibility, and a moat no horizontal builder can replicate. The architecture supports it today; the relationships get built once the product is live.
The toolkit — wedge products that fund the platform
Most platforms add features to keep paying customers. Stack's tools are built to acquire new ones. Every wedge product is shippable as a standalone SaaS with its own positioning, its own pricing, its own audience — and every one carries the agent into the platform when they're ready for more.
vs. Cloud CMA · $32–50/mo · static PDFs
Paste an address, get a shareable interactive web CMA in 20 seconds. Chat to refine — exclude a comp, test a higher price, rerun for buyer-side. AI writes the narrative and the agent's talking points. The seller clicks through it; the agent sounds prepared.
CRM consultant for Follow Up Boss
Plug in a FUB key, get a free full-account audit — dead leads, stale tags, source-ROI, missed follow-ups. Upgrade to credits, the AI builds the action plans, drips, and smart lists, and quietly stays accountable for the next 90 days.
vs. nothing — agents read 80-page PDFs by hand
Upload the inspection report. Get a one-page agent-perspective summary: must-fix vs. cosmetic, estimated repair costs, talking points for the listing agent, redlines for the addendum. The buyer's agent's most-asked-for tool, finally.
vs. nothing — low appraisals kill deals worth $20K commission
Drop in a low appraisal. AI cross-references comps the appraiser used vs. comps that should have been used, drafts the rebuttal letter, generates the supporting brief. Saves deals that otherwise die in negotiation.
vs. Highnote, Coffee & Contracts · $20–40/mo
Branded 4-tab interactive pitch from a single address + voice profile. Sixty seconds to a presentation that took three hours in PowerPoint. Already shipped in the Realty Stack open-source plugin.
vs. nothing — buyer consults rely on the agent's memory
Buyer name + criteria → AI synthesizes neighborhood profiles, school data, market trajectory, and recommended search areas into a branded buyer-presentation packet. The buyer takes Stack home; Stack stays in their head.
Each tool is a tweet that writes itself. Cloud CMA but agentic. CRM consultant on tap. Inspection PDFs translated into agent-language in 20 seconds. The build-in-public flywheel funds the audience that funds the platform. None of this requires winning the platform battle on day one. Stack wins wedge by wedge.
Built by an agent
"Every tool I'm supposed to use as a working agent was built by people who've never closed a deal. Stack is the first one built from the other side of the table."
Stack isn't a tech founder's bet on real estate. It's an agent's bet on the moment AI capability finally caught up to what the job actually needs. The product roadmap is shaped by a decade of using broken tools and watching coworkers fail at the same workflows for the same structural reasons. Agent psychology — what they'll actually adopt, what they'll abandon, what they'll champion — is the design system underneath the design system.
What's already running
Stack is being built on top of foundation work that's already in production. Code, with tests, deployed on real infrastructure, charging real money where the surfaces are live.
Public Claude Code plugin — 8 shipped skills, v0.0.5 in 3 days, 58 commits, MIT-licensed. CMA, listing presentation, buyer presentation, voice draft, brand kit. The lead-gen wedge for the paid stack.
Phase 1 prototype done, Phase 2 in progress. The /audit skill runs end-to-end against any Follow Up Boss account today. Per-customer Supabase Postgres with pgvector. 21-skill registry, billing model researched.
Production infrastructure for spawning multiple AI tools on shared rails. Resend live, Stripe webhooks live for invoice.paid + checkout.session.completed, Supabase prod, CF Workers deployed, admin cockpit at tools.holdengr.com.
Hurdle 1 foundation slice shipped — schema → renderer pipeline working, locked compliance slot injection guarded by test invariant. Plans 3, 4, 4.5, and 5A complete: AI codegen + two-stage compliance pipeline + multi-turn editing, verified against real Anthropic API.
Stack opens with a focused early-access cohort. Direct line into the roadmap. Hands-on onboarding from Holden. The list closes when it closes.