NL-GOV-MCP

Dutch public-sector data is scattered across many sources that do not natively work together. CBS does not know what Tweede Kamer publishes. BAG does not know what DUO knows. Rechtspraak is disconnected from Rijksbegroting.

NL-GOV-MCP connects what the Dutch government has not connected itself: one interface, many sources, one question, one answer — with provenance.

It is an open-source Model Context Protocol server that lets AI assistants search, combine, and return data from Dutch public-sector sources. Built by WAINUT, a one-stop AI shop in the Netherlands (AI Recruitment, AI Consulting & Implementation, AI & Data Training).

What can you do with this?

Ask in plain Dutch or English. The server routes to the right sources, retrieves data, and returns structured results with source traceability.

Examples:

• "Hoeveel sociale huurwoningen zijn er gebouwd in Rotterdam sinds 2020?" → combines relevant housing/statistics sources

• "Wat heeft de Tweede Kamer besloten over stikstof afgelopen maand?" → parliamentary search with temporal parsing

• "Welke basisscholen zijn er in Tilburg?" → real per-school records from DUO (address, denomination, BRIN)

• "Welke middelbare school scoort het best?" → DUO exam results per school location (pass rate, average marks)

• "Wat besteedt provincie Overijssel aan?" → TenderNed procurement notices and awards

• "Hoe stemde Tilburg bij de Tweede Kamerverkiezingen?" → Kiesraad results per party

• "Toon alle rechtspraak over huurrecht dit jaar" → Rechtspraak search with date-aware mapping

• "Wat is de luchtkwaliteit in Utrecht?" → live Luchtmeetnet measurements from that city’s own stations

• "Geef me de rijksbegroting voor onderwijs" → Rijksbegroting search + chapter navigation

Related MCP server: Qdrant Neo4j Crawl4AI MCP Server

How is this different from data.overheid.nl?

data.overheid.nl is primarily a catalog that tells you where data lives.

NL-GOV-MCP actively retrieves and normalizes data across many sources, can combine cross-source results, and returns a consistent MCP response contract ready for assistants and automations.

Sources (44 connectors, 70 tools)

Key features

Consistent response contract

Every tool returns the same shape:

• summary

• records[]

• provenance

• optional access_note

• optional failures[]

• optional pagination (offset, limit, total, has_more)

• optional verbose (request timings, connector health snapshots)

PDF text extraction

Most Dutch government "data" is text inside a PDF. Tools that reach a PDF resource extract its text layer instead of handing back a link only:

• tweede_kamer_document_get with include_text: true returns the text of a Kamerstuk PDF (text_preview_source: "pdf_text_layer", plus page count)

• tenderned_aanbesteding_get with include_text: true returns the text of the official tender notice PDF

Extraction is capped (max_chars, default 12 000) and fails typed rather than hard: a scan without OCR reports no_text_layer, an encrypted file encrypted, an HTML error page not_a_pdf.

Shared geo primitive

Bbox-driven sources (BRON verkeersongevallen, ruimtelijke plannen, BRP gewaspercelen) accept a gemeente name and resolve it to an RD (EPSG:28992) bbox through one shared implementation (src/utils/geo.ts), with a single extent validation and a consistent access_note when a name cannot be resolved.

Built-in resilience (zero-config)

No setup required — the following run automatically in-process:

• Per-connector circuit breaker (auto-disables after repeated failures, probes for recovery). A source whose primary endpoint is expected to fail over gives its fallback its own connector name, so a degraded primary cannot lock out the path that still works (see Luchtmeetnet).

• Per-connector concurrency limiter (default 3 in-flight, overflow queued with timeout)

• In-process HTTP response cache with hardcoded TTL per source category

• Per-connector health counters (exposed via /health/sources on SSE transport)

Graceful error handling

Typed errors:

• timeout

• http_error

• rate_limited

• malformed_response

• not_configured

• circuit_open

• unexpected

This lets assistants respond meaningfully instead of failing hard.

Structured output & debug modes

• outputFormat: json (default), csv, geojson, markdown_table

• offset / limit: pagination with metadata

• dryRun: shows planned API calls without executing them

• verbose: adds request timings, fallback steps, and connector health snapshots

Available on nl_gov_ask and major individual tools: cbs_tables_search, cbs_observations, data_overheid_datasets_search, duo_datasets_search, tweede_kamer_documents, tweede_kamer_search, officiele_bekendmakingen_search, rijksoverheid_search, rijksbegroting_search, overheid_api_register_search.

CBS trend enrichment

• cbs_observations injects lightweight trend fields when the result shape clearly supports it:

  • previous_period

  • previous_value

  • delta

  • delta_pct

• This only activates when there is a single clear period dimension and one numeric measure, so it stays inert on ambiguous wide tables.

Smart routing + temporal parsing

• nl_gov_ask routes by intent, and can run multi-source queries in parallel.

• Natural date expressions in NL/EN are currently parsed in nl_gov_ask and mapped to source filters (vorige week, sinds 2020, between 2018 and 2022, etc.).

• Temporal parsing is resolved server-side with a real reference timestamp, cross-platform via Node runtime APIs (Windows/macOS/Linux).

• Default timezone: Europe/Amsterdam.

• Override options for nl_gov_ask:

  • tool input: timezone

  • tool input: reference_now

  • environment: NL_GOV_TIMEZONE

  • config: config/default.json → temporal.defaultTimeZone

Cross-reference linking

Post-processing adds related_links[] when records share key identifiers (e.g. ECLI, BWBR, municipality codes), and can enrich legal references with direct links to wetten.overheid.nl.

Quick start

Requires Node.js >= 22.

npm ci
npm run build
npm run dev                    # start stdio server (for Claude Desktop / Claude Code)
npm run dev:sse                # SSE/HTTP server on port 3333
npm run dev:streamable-http    # Streamable HTTP server on port 3333 (MCP spec 2025-03-26)

To verify your setup:

npm run check        # type-check without emitting
npm test             # unit tests
npm run test:questions  # integration test suite (offline fixtures)
npm run test:live    # integration test suite (live API calls)

Configuration

Transport modes

Three transport modes are supported. All expose the same 70 tools.

stdio (Claude Desktop, Claude Code)

npm run dev     # development
npm run start   # production

SSE/HTTP (Open WebUI, legacy MCP clients)

npm run dev:sse    # development
npm run start:sse  # production

Streamable HTTP (MCP spec 2025-03-26)

npm run dev:streamable-http    # development
npm run start:streamable-http  # production

Session management uses the mcp-session-id header.

Selecting transport via environment

Instead of CLI flags, you can set MCP_TRANSPORT:

MCP_TRANSPORT=sse node dist/src/index.js
MCP_TRANSPORT=streamable-http node dist/src/index.js

Docker

docker build -f docker/Dockerfile -t nl-gov-mcp .
docker run --rm -p 3333:3333 \
  -e KNMI_API_KEY=your-key \
  -e OVERHEID_API_KEY=your-key \
  -e BAG_API_KEY=your-key \
  -e DSO_API_KEY=your-key \
  nl-gov-mcp

Claude Desktop integration

Build the project, then add an entry to your Claude Desktop config.

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

• Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "nl-gov-mcp": {
      "command": "node",
      "args": ["/absolute/path/to/NL-GOV-MCP/dist/src/index.js"],
      "env": {
        "OVERHEID_API_KEY": "...",
        "KNMI_API_KEY": "...",
        "BAG_API_KEY": "...",
        "DSO_API_KEY": "...",
        "NL_GOV_TIMEZONE": "Europe/Amsterdam"
      }
    }
  }
}

Restart Claude Desktop after saving.

Environment variables

Running behind a proxy

Node's built-in fetch — which every connector uses — ignores HTTP_PROXY / HTTPS_PROXY. On a network that only allows outbound traffic through a proxy, requests therefore go out directly and individual sources start failing with confusing statuses (403, 406, timeouts) while curl to the same URL from the same machine succeeds, because curl does honour those variables.

Start the server with Node's proxy support enabled:

NODE_USE_ENV_PROXY=1 HTTPS_PROXY=http://proxy.internal:3128 npm run start

Node ≥ 22 prints an "experimental" warning for this flag; it works. Symptom to recognise: some sources work and others do not, with no pattern in the code — that is an egress problem, not a connector problem.

Source-specific details

Tweede Kamer document retrieval

• tweede_kamer_documents stays lean and returns search/discovery metadata.

• tweede_kamer_document_get can optionally:

  • resolve the underlying resource URL / file metadata

  • include a capped text preview for text-like resources

  • extract the text layer of PDF resources (include_text: true), reported as text_preview_source: "pdf_text_layer" with resource_pages

• nl_gov_ask may automatically deepen the top Tweede Kamer match when the user explicitly asks for content/summary rather than only discovery.

Rechtspraak details

rechtspraak_search_ecli mirrors the official frontend search backend (/api/zoek) instead of the legacy open-data feed.

Uses structured parameters instead of natural-language parsing:

• sort: relevance (default), date_newest (publication date desc), ruling_newest (ruling date desc)

• date_filter: week, month, year, last_year (maps to Rechtspraak facet filters)

The LLM interprets user intent and maps it to these parameters. A lightweight server-side query rewriter strips residual question framing as a safety net.

Responses include facet-driven context in access_note when filters are applied.

Tuchtrecht is a separate source. Disciplinary rulings against doctors, lawyers, notaries, accountants, vets and bailiffs are published on tuchtrecht.overheid.nl, not on Rechtspraak.nl. Use tuchtrecht_search for those; nl_gov_ask routes disciplinary questions there before it considers Rechtspraak.

TenderNed details

tenderned_aanbestedingen_search sends only parameters that are verified to filter server-side (search, typeOpdracht, procedure, publicatieDatumVanaf, publicatieDatumTot, page, size). The upstream silently ignores unknown parameters, so an unsupported filter would look applied while returning everything — hence the deliberately small parameter surface. Page size is capped at 100 by the API; use page for more.

DUO per-school data

duo_schools and duo_exam_results query the CKAN datastore (the rows), not the dataset catalogue:

• duo_schools returns individual schools/institutions per sector (po, vo, mbo, ho) with address, BRIN/instellingscode, denomination and website. municipality, place and postcode filter exactly (case-insensitive input); name is a free-text search.

• duo_exam_results returns pass rates and average exam marks per school location, filterable by year, municipality and education type, with sortByScore for "which school scores best". Coverage: school years 2013–2017 — the last per-location exam dataset DUO publishes machine-readably; a year outside that range returns 0 records with an explanation in access_note rather than a validation error.

Elections (Kiesraad)

verkiezingsuitslagen_search accepts an election code (TK20251029), an election kind (TK, gemeenteraad, Europees Parlement) or nothing at all (most recent election). gebied drills down to a province or municipality; unknown areas fall back to the national result with an explanatory access_note. Use list_elections: true for the available elections.

Documentation

See:

• docs/ARCHITECTURE.md — technical internals, layer diagram, request lifecycle, resilience stack

• docs/SOURCES.md — endpoint details per connector

• docs/TOOLS.md — full tool catalog with behavior notes

• docs/BACKLOG-SOURCES.md — planned integrations

Contributing

PRs are welcome — bug fixes, new source connectors, or improvements to existing ones.

See CONTRIBUTING.md for setup, workflow, and a step-by-step guide for adding a new source connector.

License

This project is licensed under the Apache License 2.0. See NOTICE for required attribution.

WAINUT and NL-GOV-MCP are trademarks of WAINUT B.V. The Apache License 2.0 does not grant permission to use these names, trademarks, or branding to imply endorsement of derivative works. Forks and derivative works must retain the NOTICE file as required by the license.

About WAINUT — WAINUT is your one-stop AI shop in the Netherlands. We help organizations adopt AI and build an AI-enabled workforce — from recruiting the right talent, to implementing the right tools, to training teams that actually use them.

Exploring AI for your organization? → wainut.ai — Unleash Your Potential.