Sales Insights MCP (POC)

A proof-of-concept MCP server for an internal presentation: a marketing/sales user asks their AI assistant a natural-language question about daily sales, and gets back an interactive chart rendered inline in the chat, via the MCP Apps extension (SEP-1865) — with a clean markdown fallback for hosts that don't support MCP Apps.

For component breakdown, request flow diagrams, configuration reference, and a roadmap to a real enterprise deployment (auth, real data, governance), see ARCHITECTURE.md.

All data is mocked (deterministic, generated at server start). There is no real API, database, or auth — this is a demo, not a product.

Tools

Every tool returns both: a markdown summary/table in content (what any MCP host shows) and an interactive widget via _meta.ui.resourceUri (what a host with MCP Apps support renders instead). Widgets are fully interactive — their filter controls call back into the server (app.callServerTool) for fresh data without leaving the chat.

Related MCP server: Sales Dashboard

Project layout

server/
  mockData.ts   deterministic mock dataset + per-tool aggregation
  format.ts     markdown fallback formatters
  server.ts     registers the 4 tools + their ui:// resources
  main.ts       entry point (stdio or Streamable HTTP)
ui/
  shared/       theme, chart components (SVG bar/line), hooks
  apps/*/       one React entry per widget
  *.html        vite entry points, bundled to a single self-contained HTML file

Setup

npm install
npm run build      # bundles the 4 widgets into dist/ui/*.html

Running

npm run serve         # Streamable HTTP on http://localhost:3001/mcp
npm run serve:stdio   # stdio transport

Testing with Claude Desktop (stdio, no tunnel needed)

Add to your Claude Desktop MCP config:

{
  "mcpServers": {
    "sales-insights": {
      "command": "npx",
      "args": ["tsx", "server/main.ts", "--stdio"],
      "cwd": "D:\\Projects\\IDEA-PROJECTS\\income-mcp"
    }
  }
}

Restart Claude Desktop, then ask e.g. "show me the daily sales trend for the last 90 days" or "who's leading the sales leaderboard".

Testing with the ext-apps basic-host (fastest iteration loop)

git clone https://github.com/modelcontextprotocol/ext-apps.git
cd ext-apps/examples/basic-host
npm install
SERVERS='["http://localhost:3001/mcp"]' npm start

Then open http://localhost:8080 (with npm run serve running in this repo).

Testing with Claude (web) / ChatGPT / Copilot (remote HTTP connector)

These hosts need a reachable HTTPS URL. For local development, tunnel the Streamable HTTP server:

npm run serve
npx cloudflared tunnel --url http://localhost:3001

Add the resulting URL (https://<random>.trycloudflare.com/mcp) as a custom connector in the host. Claude Desktop/web and VS Code Copilot currently render MCP Apps widgets; hosts without MCP Apps support (or where the feature isn't enabled) will fall back to the markdown table/summary automatically — that's the graceful-degradation path this POC is built to demonstrate.

AI Assistant (chat UI, port 3002)

A second, separate piece of this repo: a standalone chat UI that talks to an LLM, calls tools on any number of MCP servers (this one included), and renders their widgets inline — with a pluggable LLM backend (Anthropic, Gemini, AWS Bedrock, or a local Ollama model, switchable via one env var) and support for adding arbitrary MCP servers at runtime. Independent of everything above — runs on its own port, doesn't require npm run serve.

npm run build && npm run assistant   # http://127.0.0.1:3002

Full setup reference — switching providers, running a model locally with Docker to sidestep API rate limits, adding MCP servers, tunneling, troubleshooting — is in ASSISTANT.md.

Notes for the presentation

• Cross-host compatibility is aspirational, not guaranteed — MCP Apps (SEP-1865) shipped as an official MCP extension in January 2026 and client support is still rolling out unevenly. Claude Desktop currently has the most mature support; treat the other hosts as "same server, same tools, same fallback text" even if the interactive widget doesn't render there yet.

• All four tools are interactive widgets (per POC scope), but the fallback text is what proves the server works everywhere regardless of UI support.

• The mock dataset is seeded, so numbers are stable across runs/demos.