Skip to main content
Glama
henriv
by henriv

ehr-mcp

A small MCP server (streamable HTTP) over the Estonian Building Register (Ehitisregister). It answers, for a building: what is it, where is it, and does it have a valid permit — using only the public, unauthenticated EHR endpoints. Geometry is never returned.

Tools

Building and address

  • 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.

  • address_lookup — input query (free-text address), optional limit (default 8). Resolves an address to candidates, each with a katastritunnus and building ehrCode, via the In-ADS gazetteer. Feed the ehrCode into ehr_building_data. See docs/address-lookup.md.

Permit and proceeding checks — see docs/permits.md for the upstream contract, the classification rules and the verified endpoint list.

  • ehr_permit_check ⭐ — the main tool. One call answers all four permit questions (kasutusluba, kasutusteatis, ehitusluba, ehitusteatis). Each returns a status, not a boolean: olemas | puudub | kehtetu | menetluses, with the documents that justify it (number, date, state, authority, legal framework).

  • ehr_document_list — the low-level building block: the building's public document list, each entry classified and marked with its legal framework.

  • ehr_proceeding_check — proceeding states by ehr_kood or document_id. Derived from the public documentState field, and says so in every response.

  • ehr_registry_part_check — land-register (kinnistusraamat) registriosa number. EHR does not hold this field, so the answer is always registriosa_number: null plus the cadastral number to continue with at RIK.

  • ehr_full_check — all six checks in one answer, by ehr_kood or aadress.

  • Upstream: https://livekluster.ehr.ee — the building, document and classifier APIs, all public, no auth. Configurable via EHR_ROOT_URL (prelivekluster / testkluster / devkluster).

  • Output budget: every compact result stays under ~2 KB (ehr_permit_check ≈1.5 KB for a 20-document building), so it is cheap to load into model context. taielik: true opts into the full data. See docs/upstream.md for the buildingData 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 seven tools. Then call one 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.

The permit check on Roseni tn 7, Tallinn:

npx @modelcontextprotocol/inspector --cli --transport http --server-url http://localhost:3000/mcp --method tools/call --tool-name ehr_permit_check --tool-arg ehr_kood=120542346

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 .envMCP_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 buildnode 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; EHR_ROOT_URL defaults to the same host and only needs setting to target a non-live cluster.

  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>"

Known limitations

These are properties of the public EHR API, not of this server. They are surfaced in the tool output rather than papered over.

  1. Attached document files are closed. EHR has restricted access to public documents containing personal data. fileInfos is often null even when the document exists, and some documents answer 401 on their detail view — those are marked juurdepaas_piiratud: true rather than dropped.

  2. You cannot search for a document by its number. POST /document/v1/document/search requires authentication. Publicly, documents are reachable only through a building. Searching by permit number needs your own index or an authenticated contract.

  3. Full proceeding data requires an X-tee/TARA agreement with the Land and Spatial Agency (Maa- ja Ruumiamet). /api/proceeding/v1/** returns 401. ehr_proceeding_check therefore derives the state from the public documentState field and labels the result as derived in every response.

  4. The registriosa number requires a separate RIK integration. EHR holds no land-register reference — verified absent from v3/buildingData, the document detail view and the archive records. ehr_registry_part_check returns null plus the cadastral number to query the land register with.

  5. Pre-2003 buildings often carry only a "Hooneregistri ehitise teatis" (DOTY 91511). That is not a kasutusluba. Such buildings get a warning, not a false positive.

Acceptance checklist

  • Inspector tools/list returns all seven tools

  • tools/call with a valid EHR code returns trimmed JSON < 2 KB, no kujud key

  • Invalid EHR code returns a friendly not-found message

  • ehr_permit_check on 120542346 → ehitusluba: olemas (12229, 12291), kasutusluba: olemas (12329, 12391), applications 11229/11329 not counted as permits

  • ehr_permit_check on 101018690 → all four puudub + pre-2003 warning

  • ehr_registry_part_check never returns a number

  • ehr_proceeding_check always marks its answer as derived

  • Request without bearer token (when MCP_TOKEN set) → 401

  • From claude.ai, "Mis hoone on EHR koodiga ?" triggers the tool (verify after deploying + connecting)

Project layout

src/
  config.ts        env-derived config (PORT, EHR_BASE_URL, EHR_ROOT_URL, MCP_TOKEN)
  index.ts         Express app: /healthz + stateless POST /mcp
  mcp.ts           McpServer + all seven tool registrations
  ehr/
    client.ts      getBuildingData() with timeout + typed errors
    trim.ts        trimBuildingData() / fullBuildingData(), drops geometry
    types.ts       loose buildingData response types
    http.ts        fetchJson() timeout + backoff retry, bounded concurrency
    classifier.ts  DOTY classification by name, 24 h cache, offline fallback
    documents.ts   document list (15 min cache) + document detail
    permits.ts     pure status derivation + warnings (no I/O)
    checks.ts      orchestration for the five check tools
  inads/
    client.ts      lookupAddress() -> In-ADS gazetteer, validation + shortcut
    parse.ts       parseCandidates() -> group by adr_id, cadastral + EHR code
    types.ts       gazetteer row + candidate types
docs/upstream.md        ehr_building_data upstream contract + trim map
docs/address-lookup.md  address_lookup upstream contract + parsing
docs/permits.md         permit-check endpoints, DOTY rules, status derivation
test/              Vitest: trim, client, address, http, classifier, permits,
                   checks — plus live-captured response fixtures

Latest Blog Posts

MCP directory API

We provide all the information about MCP servers via our MCP API.

curl -X GET 'https://glama.ai/api/mcp/v1/servers/henriv/ehr-mcp'

If you have feedback or need assistance with the MCP directory API, please join our Discord server