Skip to content

Releases: DaveGold/mcp-metadata-demo

v1.1.0 — minimal ablation tier + render_table rework

Choose a tag to compare

@DaveGold DaveGold released this 28 Aug 13:18
f6a282d

Second release of the Missing Layer demo. Everything here is backward-compatible with the v1.0.0 rich API.

Added

  • Minimal ablation endpoint (mcpMinimal) — a metadata-stripped twin of get_building_profile: one-sentence description, no schema, no curated alerts[]. Same data, no metadata layer, so the paper's thesis can be tested by contrast.
  • Apps on the minimal tier — the minimal endpoint exposes the same tool set as rich (get_building_profile + render_chart / render_table / render_map), every tool stripped to minimal instructions. The only variable between the two endpoints is the metadata itself.
  • Dual hosted endpoints, documented with a runnable A/B prompt and deep links to the metadata in source.

Changed

  • render_table rework — migrated to TanStack Table v9 stable, fixed pagination:false (renders all rows) and global search, refactored footer aggregation.
  • README restructured for a conference-first read: paper link → thesis → "try it live" → why → depth.

Endpoints

  • rich: https://europe-west4-mcp-metadata-demo.cloudfunctions.net/mcp
  • minimal: https://europe-west4-mcp-metadata-demo.cloudfunctions.net/mcpMinimal

📄 Companion to The Missing Layer.

v1.0.0 — mcp-metadata-demo

Choose a tag to compare

@DaveGold DaveGold released this 29 May 13:25

First public release — companion demo to the paper The Missing Layer.

A working demonstration of a rich-metadata strategy for AI tooling, applied at two levels in one MCP server:

  • Rich-domain toolget_building_profile fuses Dutch open building data (BAG via PDOK + EP-Online energy labels) into a structured profile with a curated alerts[] array, so an agent reasons about the domain without external priming.
  • Self-describing MCP appsrender_chart, render_table, render_map apply the same metadata strategy to UI configuration; the schemas tell the agent which chart type fits the data, which cell formatters apply, and which marker types to use.

What's inside

  • 5 tools: get_building_profile, render_chart, render_table, render_map, fetch_image
  • Three transports from one createServer() factory: stdio, local HTTP, Firebase Cloud Function
  • MCP Apps UI pipeline: Angular + Vite single-file bundles, Tailwind, Chart.js / Leaflet / TanStack Table
  • Hosted public demo endpoint (no auth, rate-limited)

Try it

Add the hosted endpoint to your MCP client:

{ "mcpServers": { "metadata-demo": { "url": "https://mcp-jtc4p3l6nq-ez.a.run.app" } } }

Or run locally — see the README for stdio setup.