ehr-mcp

ehr-mcp

MCP server that exposes the ehr_building_data tool to query the Estonian Building Register by EHR code, returning a compact summary of building info (address, technical data, usage, energy certificate, cadastral units) excluding geometry.

Category
Visit Server

README

ehr-mcp

A small MCP server (streamable HTTP) that exposes one tool backed by the Estonian Building Register (Ehitisregister) endpoint GET /v3/buildingData. It lets you ask Claude about a building by its EHR code and get back a compact, trimmed answer — address, key technical indicators, usage purpose, energy certificate and cadastral units. Geometry is never returned.

  • Tool: ehr_building_data — input ehr_kood (numeric string), optional taielik (boolean). Default returns a compact <2 KB summary; taielik: true returns every field except geometry (~6 KB compact). Geometry is never returned in either mode.
  • Upstream: https://livekluster.ehr.ee/api/building/v3/buildingData (public, no auth).
  • Output budget: trimmed result stays well under ~2 KB (≈0.5 KB typical) so it is cheap to load into model context. See docs/upstream.md for the full contract and trim decisions.

Stack

Node 20+, TypeScript (strict), Express, @modelcontextprotocol/sdk, zod, Vitest.

Develop

npm install
cp .env.example .env      # optional; leave MCP_TOKEN empty for no-auth local dev
npm run dev               # tsx watch on http://localhost:3000

Other scripts: npm run build (tsc → dist/), npm start (node dist/index.js), npm test (Vitest).

Health check: GET /healthz → { "ok": true }.

Smoke test with MCP Inspector

Start the dev server (npm run dev), then in another terminal:

npx @modelcontextprotocol/inspector --cli --transport http --server-url http://localhost:3000/mcp --method tools/list

Expect one tool, ehr_building_data. Then call it with a real EHR code (101018690 is the Tallinn sample used in the tests):

npx @modelcontextprotocol/inspector --cli --transport http --server-url http://localhost:3000/mcp --method tools/call --tool-name ehr_building_data --tool-arg ehr_kood=101018690

Expect trimmed JSON (< 2 KB, no kujud key). An unknown code such as 120896 returns a friendly EHR koodiga 120896 ehitist ei leitud. message.

If MCP_TOKEN is set, add --header "Authorization: Bearer <token>" to the Inspector commands.

Auth

Compatibility note: the claude.ai web custom-connector dialog authenticates only via OAuth (Client ID + Client Secret) — it has no field for a static bearer token. So the bearer token below cannot be used from the web connector; to use claude.ai, run the service authless (leave MCP_TOKEN unset — see Connect to Claude). The bearer token still works anywhere a custom header can be sent: Claude Code (--header), the MCP Inspector, and direct curl.

Auth on POST /mcp is a single shared bearer token — it is a secret string, not a hash or a signed token, and there is no issuance, expiry, or user model by design. /healthz is always open. If MCP_TOKEN is unset/empty, /mcp is open too (local no-auth dev); if it is set, every request must send Authorization: Bearer <token>. The comparison is constant-time (crypto.timingSafeEqual, length-checked first so it can't throw).

Generate the token once:

openssl rand -hex 32
# or
node -e "console.log(require('crypto').randomBytes(32).toString('hex'))"

The same value goes in three places:

  1. Local .env — MCP_TOKEN=... (gitignored, never committed).
  2. Render → your service → Environment → MCP_TOKEN. This is why render.yaml marks it sync: false: the value is entered in the dashboard, never stored in the repo.
  3. Wherever you send it as a header — Claude Code (--header "Authorization: Bearer ..."), the MCP Inspector (--header), or curl. (Not the claude.ai web connector — see the compatibility note above.)

Rotation: generate a new value, update it in Render's Environment and in the connector, and redeploy. The old token stops working the moment the new value is live — there is no grace window or revocation list to manage.

Verify auth against the deployed service (replace host + token):

# 401 without the header
curl -s -o /dev/null -w "%{http_code}\n" -X POST https://ehr-mcp.onrender.com/mcp \
  -H "content-type: application/json" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

# 200 with it
curl -s -o /dev/null -w "%{http_code}\n" -X POST https://ehr-mcp.onrender.com/mcp \
  -H "content-type: application/json" \
  -H "Authorization: Bearer <token>" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

Deploy to Render

The repo includes render.yaml (a web service: npm ci && npm run build → node dist/index.js, health check /healthz).

  1. Push this repo to GitHub.
  2. In the Render dashboard: New → Web Service (or Blueprint to pick up render.yaml directly) and connect the GitHub repo.
  3. Set the MCP_TOKEN environment variable in the dashboard (see Auth). EHR_BASE_URL is supplied by render.yaml.
  4. Deploy. The service will be at https://<service>.onrender.com, with the MCP endpoint at https://<service>.onrender.com/mcp.

Free-tier cold start: the free plan spins the service down after ~15 min idle; the next request then waits ~30–50 s while it wakes. The first Inspector or Claude call after idle may look like a timeout — retry. Render's Starter tier keeps the service always-on and removes the cold start.

Connect to Claude

claude.ai (custom connector): Settings → Connectors → Add custom connector → URL https://<service>.onrender.com/mcp. Leave the OAuth Client ID / Client Secret fields empty — the web connector only supports OAuth, not a static bearer token, so the service must be running authless (MCP_TOKEN unset in Render). Once connected, ask in Estonian, e.g. "Mis hoone on EHR koodiga 101018690?" — Claude will call ehr_building_data and answer with the address and key indicators.

Running authless means anyone with the URL can call the tool. That's low-risk here — it only reads public, read-only Ehitisregister data — but it does mean your Render compute is open. If you need it locked down, use Claude Code with the bearer token instead of the web connector.

Claude Code (CLI):

claude mcp add --transport http ehr https://<service>.onrender.com/mcp

If auth is enabled, add the header:

claude mcp add --transport http ehr https://<service>.onrender.com/mcp --header "Authorization: Bearer <token>"

Acceptance checklist

  • [x] Inspector tools/list returns ehr_building_data
  • [x] tools/call with a valid EHR code returns trimmed JSON < 2 KB, no kujud key
  • [x] Invalid EHR code returns a friendly not-found message
  • [x] Request without bearer token (when MCP_TOKEN set) → 401
  • [ ] From claude.ai, "Mis hoone on EHR koodiga <code>?" triggers the tool (verify after deploying + connecting)

Project layout

src/
  config.ts        env-derived config (PORT, EHR_BASE_URL, MCP_TOKEN)
  index.ts         Express app: /healthz + stateless POST /mcp
  mcp.ts           McpServer + ehr_building_data tool
  ehr/
    client.ts      getBuildingData() with timeout + typed errors
    trim.ts        trimBuildingData() -> <2 KB, drops geometry
    types.ts       loose response types
docs/upstream.md   upstream contract, params, response shape, trim map
test/              Vitest: trim + client, plus the raw response fixture

Recommended Servers

playwright-mcp

playwright-mcp

A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.

Official
Featured
TypeScript
Magic Component Platform (MCP)

Magic Component Platform (MCP)

An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.

Official
Featured
Local
TypeScript
Audiense Insights MCP Server

Audiense Insights MCP Server

Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.

Official
Featured
Local
TypeScript
VeyraX MCP

VeyraX MCP

Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.

Official
Featured
Local
graphlit-mcp-server

graphlit-mcp-server

The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.

Official
Featured
TypeScript
Kagi MCP Server

Kagi MCP Server

An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.

Official
Featured
Python
E2B

E2B

Using MCP to run code via e2b.

Official
Featured
Neon Database

Neon Database

MCP server for interacting with Neon Management API and databases

Official
Featured
Exa Search

Exa Search

A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.

Official
Featured
Qdrant Server

Qdrant Server

This repository is an example of how to create a MCP server for Qdrant, a vector search engine.

Official
Featured