case study

Perch

Perch is a living civic-data platform: a set of 3D city maps you can read at a glance, and an open MCP that serves the same data to AI agents. It started as a single map of Austin and is now a multi-city system — Austin, with New York, Washington DC, and Seattle live in beta — plus a connector that lets assistants like Claude query civic data across 23 US metros.

Perch is my own project, and a long time coming: I've wanted to build a real-life SimCity ever since the game consumed my childhood. I came up through web design — fluent in HTML and CSS, lighter on JavaScript — so a full 3D, data-driven app stayed out of reach for years. Claude Code is what finally made it possible. I design and build it end to end, using the AI-assisted workflow (Claude Code and GitHub) I now use daily as a product developer. The immediate spark was Corpus Day Zero, a drought dashboard I built for the Nueces Basin — Perch is that idea scaled to a whole city, and now to many.

Perch — a 3D isometric civic dashboard of downtown Austin along Lady Bird Lake

WHAT IT IS

Perch is two things built from one source of truth: living 3D maps for US cities, and an open MCP that serves the same civic data to AI agents. Each map renders a city as a single continuous isometric model — 3D buildings, terrain, day-to-night light — with toggleable layers for 311, crime, crashes, permits, weather, transit, and more, laid onto the streets and buildings where they actually happen.

Austin is the flagship, fully enriched — a left rail of data layers, a City / State toggle for level of government, live transit, and city-specific touches. New York, Washington DC, and Seattle are live in beta. Rather than a wall of charts, the data lives in context: the map is the dashboard.

THE PROBLEM

A city already publishes an enormous amount of data — weather, traffic, 311 requests, water, power, permits, crime — but it lives in dozens of separate agency dashboards and open-data portals, each with its own map, login, and learning curve. Even a motivated resident rarely stitches it together.

Across cities it's worse: every metro runs a different system, with different dataset names and different ways to ask. What's a "311 request" in one city is something else entirely in the next. Perch starts from the opposite premise — one place, one way to ask, everything in context — and makes that true for people and for software at the same time.

ONE CONTRACT, TWO PRODUCTS

Perch hinges on a single idea: describe each city once, and let that description drive everything. Every city is one declarative contract — a file that names its datasets, says where each one lives, and maps its fields to a shared vocabulary (category, date, status, area, location). The map engine and the MCP both read the same contracts; neither hard-codes a single city. Adding a city is writing a contract, not editing the app.

  • The engine carries no city knowledge. The map reads which city it's showing from the URL and builds itself from that city's contract — so Austin's bespoke touches and a new beta city run on exactly the same code.
  • Four adapters, one question. Cities don't publish data the same way — some run Socrata, others ArcGIS, CKAN, or Carto. A small set of adapters translates one canonical query into each backend's native dialect, so the same question works everywhere.
  • A coverage map, not a lowest common denominator. Where a city doesn't publish a dataset, the system says so plainly instead of faking it. The differences between cities are information worth keeping, not noise to flatten.

DESIGN & BUILD

A few of the decisions behind it:

  • An isometric 3D model instead of charts. A grid of graphs would have been faster to build, but data reads faster when it sits on the actual street or building it describes. The 3D city is the navigation and the visualization at once.
  • One unified map with toggleable layers, rather than a separate page per dataset. Layers stack and combine, so you can see traffic, weather, and 311 in the same glance instead of tabbing between tools.
  • A soft, illustrative art direction — rounded forms and calm color — over the dark "mission-control" look most civic dashboards default to. The goal was something a resident would open for pleasure, not just utility.
  • Built solo with an AI-assisted workflow. A React and Vite front end on Mapbox GL (3D buildings, terrain, and day-to-night light), with the map and the MCP sharing the same query core — developed with Claude Code and versioned on GitHub.

WRITING THE AI

A lot of Perch is content design, not chart-plotting. Several of its surfaces are generated by Claude, and each only works if it speaks in a single, trustworthy civic voice. For each, I wrote the system instructions, the voice-and-tone rules, and the reference examples that steer the model — then graded the output against them and revised the instructions until it held the voice.

  • Ask Perch, a conversational assistant. I wrote the system prompt, voice rules, and reference answers that shape how a Claude-based assistant fields residents' questions about their city, then built a set of test questions across categories and ran them through repeatedly, tightening the instructions based on what came back.
  • Perch Weekly, a generated newspaper. Claude writes a short civic paper each week from the data Perch collects. I defined the sections, the editorial voice, and the rules for turning raw data into accurate, readable stories — attribute every claim, separate what the data proves from what it merely suggests, no hype. Early issues fixated on routine 311 requests and permit filings, so I audited each one and revised the editorial rules until coverage landed on stories that actually mattered. It's been published weekly since June 2026.
  • Build proposals, written for officials. Residents sketch on the map and describe what they want built; the model writes a full proposal with cost estimates, addressed to the right council members. I wrote the structure and tone so a civic audience gets something credible and specific, not boilerplate.

AN OPEN MCP FOR AI AGENTS

The same contracts that draw the maps also feed an open MCP server, so the civic data is as usable by an AI assistant as by a person. Instead of learning each city's schema, an assistant like Claude asks one set of questions — 311, crime, crashes, permits, and more — and gets consistent answers back across every metro, each with its source and freshness attached.

It covers 23 US metros through ten read-only tools: discovery for cities and datasets, a canonical query with filters for category, date, status, and area, a raw-query escape hatch, and live national feeds for weather, wildfires, wildlife, population, and demographics. Four adapters (Socrata, ArcGIS, CKAN, and Carto) translate that one query into each city's native system. It runs on a Cloudflare Worker and also exposes a plain REST API, with a companion dashboard for comparing cities side by side.

It's read-only by construction — it never writes anything and returns only already-public open-government data. Anyone can add it to Claude under Settings → Connectors, with no account and no key, and start asking about their own city. It's the same principle as the rest of Perch, extended to agents: give an AI trustworthy, well-sourced data to reason over, without ever letting it change the record.

Perch's Cross-City Compare dashboard: a 'Crime · monthly trend' line chart comparing several US cities, each chip labeled by its open-data backend (Socrata, ArcGIS, CKAN, Carto)

OUTCOME

Perch was selected to present at Mapbox BUILD 2026, the company's developer conference — recognition from the platform it's built on that the approach is worth showing to other builders. It's now a multi-city system with a public MCP in beta, and development is ongoing: more cities, more layers, and live maps for the metros that are MCP-only today.