@propraven/mcp

Model Context Protocol server for PropRaven — gives Claude, ChatGPT, Cursor, and any MCP-compatible agent canonical access to 191.3M US parcels (110.0M mapped) with ownership, valuation, permits, deeds, hazard, and market data.

Hosted endpoint: https://propraven.com/mcp (Streamable HTTP, no install — docs: https://propraven.com/docs/mcp · agents + x402: https://propraven.com/docs/agents). Or run the stdio server locally (below).

Hosted (no install)

Point any HTTP-transport MCP client at https://propraven.com/mcp with your API key as a bearer header. Claude Desktop / Cursor (claude_desktop_config.json, ~/.cursor/mcp.json):

config.json

{
  "mcpServers": {
    "propraven": {
      "url": "https://propraven.com/mcp",
      "headers": { "Authorization": "Bearer pz_your_real_key_here" }
    }
  }
}

Claude Code: claude mcp add --transport http propraven https://propraven.com/mcp --header "Authorization: Bearer pz_your_real_key_here". The claude.ai web connector uses OAuth instead — add a custom connector with the same URL and sign in.

Local install in Claude Desktop

Requirements: Node ≥18, a PropRaven API key (pz_…) from https://propraven.com/settings/api-keys (free, no card).

One-time setup

bash

git clone https://github.com/jdw2111/propraven-mcp.git
cd propraven-mcp
npm install
npm run build

Wire into Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

config.json

{
  "mcpServers": {
    "propraven": {
      "command": "node",
      "args": ["/absolute/path/to/propraven-mcp/dist/index.js"],
      "env": {
        "PROPRAVEN_API_KEY": "pz_your_real_key_here"
      }
    }
  }
}

Restart Claude Desktop. The 8 PropRaven tools (parcel.lookup, parcel.search, parcel.compare, owner.pierce, hazard.score, valuation.estimate, permits.history, sales.history) appear in the tool list.

Quick test

Code

Look up parcel 37183:0012345 and tell me the owner.

Claude should pick parcel.lookup and return canonical record + owner.

Cursor + ChatGPT

Cursor: install via the MCP marketplace (search "PropRaven") once we publish, or wire manually:

jsonc

// ~/.cursor/mcp.json
{
  "mcpServers": {
    "propraven": { "command": "node", "args": ["/abs/path/dist/index.js"], "env": { "PROPRAVEN_API_KEY": "pz_..." } }
  }
}

ChatGPT: Custom GPT → Actions → OpenAPI URL https://api.propraven.com/openapi.json with API-key (Bearer) auth; native MCP-in-ChatGPT can use the hosted endpoint https://propraven.com/mcp (OAuth 2.1 + PKCE).

Tools

Tool	Purpose	Backed by
`parcel.lookup`	Resolve one parcel by composite ID, address, or APN	`GET /api/v1/parcels/{id}` or `GET /api/v1/search/full`
`parcel.search`	Filter parcels by geography + attributes	`GET /api/v1/search/full`
`parcel.compare`	Fetch reports for 2–25 parcels for agent-side ranking	`GET /api/v1/parcels/{id}/report` × N
`owner.pierce`	Resolve a name/entity to its full portfolio	`GET /api/v1/owners/{name}/portfolio`
`hazard.score`	Flood / fire / seismic / windstorm / air-quality / crime composite	`GET /api/v1/parcels/{id}/risks`
`valuation.estimate`	AVM + assessed / market values for a parcel	`GET /api/v1/parcels/{id}`
`permits.history`	Permit timeline for a parcel	`GET /api/v1/parcels/{id}/permits`
`sales.history`	Deed timeline ± UCC liens	`GET /api/v1/parcels/{id}/deeds` (+ `/report` if liens)

Tool descriptions are optimized for agent reasoning — they explicitly call out when not to use each tool, which is the biggest determinant of agent selection accuracy.

Environment

The server reads:

Variable	Default	Purpose
`PROPRAVEN_API_KEY`	(required)	Bearer token, format `pz_...`
`PROPRAVEN_BASE_URL`	`https://api.propraven.com`	API host. Use `https://propzilla.vercel.app` during the DNS cutover window.
`PROPRAVEN_TIMEOUT_MS`	`30000`	Per-request timeout in ms.

Develop / inspect

bash

# Live-reloading dev (TypeScript, no build step):
PROPRAVEN_API_KEY=pz_test... npm run dev

# Or open the MCP Inspector UI:
PROPRAVEN_API_KEY=pz_test... npm run inspect

The inspector at https://modelcontextprotocol.io/legacy/tools/inspector lets you call each tool manually and see request/response payloads.

Why this exists

Every Claude / ChatGPT / Cursor workflow that touches property data needs an authoritative parcel-lookup tool. There is no canonical property-data MCP today. PropRaven aims to be it before the first-mover window closes (~12–18 months).

License

Apache-2.0