Property App — MCP App with Prefab UI

The 4th consumer of property_core (alongside API, CLI, MCP server). A FastMCP 3.2+ server with Prefab UI 0.19+ dashboards, deployed on Fly.io.

Production: https://propertydata.fly.dev/mcp

Commands

# Run locally
uv run --extra apps property-app                                  # stdio
MCP_TRANSPORT=http uv run --extra apps property-app               # HTTP
uv run --extra apps fastmcp dev apps property_app/server.py:mcp   # dev preview in browser

# Tests
uv run --extra apps pytest tests/test_app_*.py -v

# Deploy
fly deploy --config fly.app.toml

Architecture

property_app/
├── server.py              # FastMCP instance, /health, /img proxy, lazy imports in main()
├── tools.py               # Plain tools + stamp_duty (app=True) + component_test, image_test
├── formatting.py          # fmt_gbp(), fmt_pct(), fmt_date()
└── dashboards/
    ├── comps.py           # search_comps + comps_dashboard (BarChart, Sparkline, Table, Tabs)
    ├── yield_view.py      # get_yield + yield_dashboard (Badge assessment, Cards)
    ├── rental.py          # get_rental + rental_dashboard (Dot status, Alert, Cards)
    └── listings.py        # listings_dashboard (listing Cards with status Badges)

Tool Inventory

Key Patterns

Data + Dashboard pairs

Each dashboard file has two tools calling the same _helper():

• Data tool (@mcp.tool()) → returns dict for LLM reasoning
• Dashboard tool (@mcp.tool(app=True)) → returns ToolResult with Prefab view

def _fetch_data(postcode: str) -> dict:
    """Raw helper — calls property_core, returns dict."""
    result = SomeService().method(postcode)
    return _slim(result.model_dump(mode="json"))

@mcp.tool()
def get_data(postcode: str) -> dict:
    """Data for LLM."""
    return _fetch_data(postcode)

@mcp.tool(app=True)
def data_dashboard(postcode: str):
    """Visual dashboard."""
    data = _fetch_data(postcode)
    view = Column(children=[...], gap=4)
    return ToolResult(content=f"Summary for {postcode}", structured_content=view)

Pre-populate, not forms

Always use @mcp.tool(app=True) with server-side data fetch. Do NOT use @app.ui() + CallTool — it renders empty forms that look broken.

Text fallback

Always return ToolResult(content=text_summary, structured_content=view). Without content, the LLM only sees "[Rendered Prefab UI]" and can't reason about the data.

_slim() on data tools

Strip raw, images, floorplans, epc_match from data tool returns to stay under token budget. Defined in tools.py.

Constructor vs context manager

Our dashboards use the constructor pattern: Column(children=[...]). FastMCP docs also show context manager syntax: with Column() as view:. Both work — be consistent within a file.

Annotations

All tools use readOnlyHint=True. External API tools add openWorldHint=True. Calculator tools add idempotentHint=True.

Tool descriptions

Neutral data-assembly language. Say "yield calculation" not "yield estimate". Say "comparable sales data" not "deal analysis". This prevents LLM consumers from expecting synthesis that isn't there.

Prefab Component Gotchas

Verified working in claude.ai as of 2026-04-11:

Metric has additional props not yet used: delta (change string like "+5.2%"), trend ("up"/"down"/"neutral"), trendSentiment ("positive"/"negative"/"neutral").

Sparkline also has: variant (color presets), strokeWidth, mode ("line"/"bar").

BarChart also has: stacked, horizontal, barRadius, showLegend, showTooltip, showGrid, yAxisFormat ("auto"/"compact").

All verified components: Column, Row, Grid, Heading, Text, Muted, Badge, Card, CardContent, Separator, Metric, Tabs, Tab, Table/TableHeader/TableHead/TableBody/TableRow/TableCell, BarChart, ChartSeries, Sparkline, Progress, Ring, Dot, Alert, ForEach, Input, Button, Form.

Adding a New Tool / Dashboard

1. Create _helper() function in a new file under dashboards/ — calls property_core, returns dict
2. Create @mcp.tool() data tool returning the helper's dict
3. Create @mcp.tool(app=True) dashboard returning ToolResult(content=text, structured_content=view)
4. Import the module in server.py main() so decorators register
5. Add test file tests/test_app_{name}.py — mock property_core calls, test both helper and dashboard

Documentation References

Append .md to any doc page URL to get raw markdown (useful for LLM context).

To check a specific component's props: https://prefab.prefect.io/docs/protocol/{component-name}.md

Deployment

• Fly.io app: propertydata in LHR (London)
• Dockerfile.app: Python 3.11 slim, uv sync --frozen --no-dev --extra apps
• Health check: GET /health every 30s (implemented in server.py)
• Transport: MCP_TRANSPORT=http with stateless_http=True
• VM: 512MB RAM, shared CPU, 1 machine minimum
• Image proxy: /img?url=... proxies Rightmove images through our domain (bypasses CSP for non-claude.ai hosts)

Environment Variables