Interactive Sales Dashboard — an MCP App for Claude

A polished, clickable sales-analytics dashboard that renders inline inside Claude (Cowork and regular chat) via the MCP Apps extension (SEP-1865, the ui:// resource pattern), served by a small Python server. Pick a date range, click a region, or pick a product and the charts and table re-query live data and re-render — a real app, not a screenshot.

• Server: FastMCP (native MCP Apps), streamable-HTTP.

• Data: stdlib sqlite3, seeded deterministically (~4,500 orders / 24 months).

• UI: one self-contained HTML doc — inline CSS/JS, inline-SVG charts (no CDN).

• Reach: a built-in Cloudflare tunnel exposes your laptop over HTTPS so Claude can reach it.

Quickstart (≈60 seconds)

You need uv and cloudflared (brew install cloudflared on macOS). Then:

uv sync                      # create the venv + install (pinned)
uv run mcp-apps seed         # build db/sales.db (deterministic; safe to re-run)
uv run mcp-apps serve        # start server + open a public HTTPS tunnel

serve prints a banner — copy the URL:

==================================================================
  CONNECTOR URL — paste into Claude → Add custom connector:
      https://<random-words>.trycloudflare.com/mcp
==================================================================

Then in Claude: Settings → Connectors → Add custom connector, paste that URL (including the trailing /mcp), enable it, and ask:

"Open the sales dashboard."

Leave serve running; press Ctrl-C to stop the server and tunnel together.

serve always opens a tunnel. If cloudflared isn't installed it stops immediately with an install hint — Claude runs in the cloud and can't reach localhost, so a public HTTPS URL is required.

The *.trycloudflare.com URL is new every run. For a stable URL on a domain you own, see Stable URL.

Related MCP server: Qlik MCP Server

Add it to Claude

1. Settings → Connectors → Add custom connector.

2. Name it (e.g. Sales Dashboard) and paste the HTTPS URL from serve, including /mcp. Save.

3. Enable it for the surface you'll use (Cowork and/or chat), then ask Claude to open the dashboard.

Use it

• "Open the sales dashboard." · "Open the sales dashboard for APAC." (pre-filtered)

• Click a region bar → the whole dashboard filters to it.

• 3M / 6M / 12M / All presets or the date pickers → time range updates.

• Click a top product → the orders table filters via a live list_orders call.

• Search / click column headers to filter and sort; click a chip × to clear.

Each click is a live MCP tool call. Open the host devtools console to watch [mcp-app] ⇒ host traffic with a matching POST /mcp in the server terminal.

How it works

Claude (host)  ──MCP over HTTPS──▶  Cloudflare tunnel  ──▶  FastMCP server  ──▶  SQLite
     │  renders ui:// resource                                    │
     └── sandboxed iframe ◀── postMessage (JSON-RPC) ─────────────┘

• Data tools (get_sales_summary, sales_by_month, sales_by_region, top_products, sales_by_rep, list_orders) run SQL and return typed JSON. get_dashboard_data bundles everything so a filter change is one round-trip.

• The app is one tool, open_sales_dashboard, linked to the UI resource ui://sales/dashboard-<hash> via _meta.ui.resourceUri. The server inlines HTML/CSS/JS into one document and embeds default data so it paints instantly.

• Interactivity flows over the SEP-1865 postMessage bridge: the iframe does the ui/initialize handshake, reports its size, and calls the data tools on each filter change.

Built against MCP Apps spec 2026-01-26, verified on FastMCP 3.4.2.

Customize / re-skin

Everything tweakable lives in the CONFIG block at the top of src/sales_dashboard/config.py: theme colours, date range & volume, regions, product catalogue & price bands, currency, server host/port. After changing data settings, re-run uv run mcp-apps seed, then restart serve and reconnect the connector.

Develop & test

uv run pytest                       # full suite (unit + in-memory integration)
uv run ruff check .                 # lint
uv run mypy src                     # type-check
# coverage gate (the suite keeps ≥90%):
uv run pytest --cov=src/sales_dashboard --cov-report=term-missing --cov-fail-under=90

Tests need no running server — integration tests drive the server in-memory via fastmcp.Client. For a bare local server (e.g. to poke at it with the MCP Inspector), run uv run server.py and point npx @modelcontextprotocol/inspector at http://localhost:8000/mcp (Streamable HTTP). Editing src/sales_dashboard/ui/* changes the content hash in the ui:// URI automatically — restart serve and reconnect so Claude fetches the new URI.

Stable URL (named tunnel)

One-time (opens a browser to log in):

cloudflared tunnel login
cloudflared tunnel create mcp-demo
cloudflared tunnel route dns mcp-demo mcp-demo.example.com

Then: uv run mcp-apps serve --named mcp-demo.example.com → stable connector URL https://mcp-demo.example.com/mcp.

Troubleshooting

Note on exposure: serve publishes your local server on a public HTTPS URL with no authentication — anyone with the URL can call the tools. The data is synthetic and the quick-tunnel URL is ephemeral, but don't point this at real data without adding auth.

Project layout

server.py                    bare local server (dev / MCP Inspector)
seed.py                      builds the SQLite DB
src/sales_dashboard/
  config.py                  ⭐ CONFIG block — theme / date range / volume / regions
  db.py                      sqlite helpers (injectable path)
  queries.py                 one SQL function per tool
  seed.py                    deterministic data generator
  server.py                  FastMCP: tools + ui:// resource; assembles the HTML
  cli.py                     `mcp-apps` CLI: seed + serve (tunnel)
  ui/                        dashboard.html / .css / bridge.js / charts.js / app.js
tests/
  unit/                      config, db, seed, queries, validation, ui-assembly, cli
  integration/               tools + app resource (in-memory FastMCP client)
docs/                        design specs, plans, screenshot

Pinned versions

MCP Apps is young and moving fast; this targets spec 2026-01-26. The bridge.js postMessage layer is isolated in one file so it can be adapted if a host's behaviour changes.

Built by dac.digital · MIT licensed.