data-gov-il-mcp

Production-grade Model Context Protocol (MCP) server for Israeli Government Open Data from data.gov.il.

This server gives MCP-compatible clients structured access to Israeli government datasets, resources, tags, organizations, and tabular records through the official CKAN API. It is written in TypeScript, validates inputs and outputs with Zod, returns JSON-first tool responses, and supports both local `stdio` and remote Streamable HTTP transports.

Highlights

Quick Start

Claude Desktop / Local Stdio

Use the published package directly:

{
  "mcpServers": {
    "data-gov-il": {
      "command": "npx",
      "args": ["-y", "data-gov-il-mcp"]
    }
  }
}

Streamable HTTP

npm install -g data-gov-il-mcp
data-gov-il-mcp-http

Then configure your MCP client:

{
  "mcpServers": {
    "data-gov-il": {
      "url": "http://localhost:3664/mcp"
    }
  }
}

Docker

docker run -p 3664:3664 ghcr.io/davidosherproceed/data-gov-il-mcp:latest

For local development:

npm install
npm run build
npm run start:stdio
# or
npm run start:http

Catalog Discovery Layer

The server ships with a committed catalog snapshot at `src/data/catalog/catalog.snapshot.json`. The snapshot is generated from data.gov.il and bundled into the build. At startup, it is validated and indexed in memory.

The discovery layer powers:

• `find_datasets` catalog-first search with live CKAN fallback.
• `list_all_datasets` instant local enumeration with optional organization filter.
• `list_organizations` instant local organization list with dataset counts.
• `list_available_tags` and `search_tags` from real CKAN tag/facet data.
• Dynamic completions for `datagov://dataset/{id}`.
• `datagov://tags` and `datagov://catalog/stats` resources.

It includes:

• Hebrew normalization, including nikud removal and final-letter normalization.
• Tokenization and trigram indexes for fuzzy matches and typo tolerance.
• Dataset, tag, and organization maps.
• Tag-to-dataset and organization-to-dataset indexes.
• Tag co-occurrence for related tag suggestions.
• Weighted ranking across exact, token, tag, organization, and fuzzy signals.

Refresh the snapshot:

npm run catalog:refresh
npm run build

A scheduled GitHub Actions workflow (`.github/workflows/catalog-refresh.yml`) refreshes the snapshot and opens a PR when catalog data changes.

More detail: `docs/catalog-discovery-layer.md`.

Tools

All successful tool responses return:

• `structuredContent`: the typed JSON object.
• `content[0].text`: the same object serialized as JSON for text-only clients.

Default Tools

Optional Tools

Elicitation in `find_datasets`

`find_datasets` can optionally expose an `interactive` parameter:

{
  "query": "תחבורה",
  "interactive": true
}

This is only registered when:

MCP_ENABLE_ELICITATION=true

When enabled, compatible clients may show a clarification form for broad searches. For example, the server can ask the user to narrow many matching datasets by publisher organization. If the client does not support Elicitation, the user declines, or the request times out, the tool falls back to normal search results.

This is disabled by default because MCP client support varies.

Resources

The server also implements resource subscriptions in a minimal standards-compliant way:

• Advertises `resources.subscribe`.
• Handles `resources/subscribe` and `resources/unsubscribe`.
• Sends `notifications/resources/updated` only for resources a client subscribed to.
• Does not poll CKAN in real time.

Prompts

Prompt arguments use MCP completions. Domain focus suggestions are curated, and organization completions are catalog-backed.

Optional MCP Client Features

These features are off by default. Enable them only when your target MCP client supports them and you want the server to expose them.

Client support differs:

• Cursor supports Elicitation, but does not currently expose Sampling.
• Claude Code supports Elicitation in recent versions.
• Claude Desktop supports many MCP features, but Elicitation support is not reliable/available.
• Sampling availability varies; the server always falls back safely.

Configuration

Copy `.env.example` to `.env`:

cp .env.example .env

Core

CKAN

HTTP Hardening

Authentication

Service Identity

Recommended Workflows

Find and Query a Dataset

1. Use `find_datasets` with natural Hebrew or English terms.

2. Use `get_dataset_info` or `list_resources` for a chosen dataset.

3. Pick a resource with `datastore_active=true`.

4. Use `search_records` with `limit=5` first to inspect fields.

5. Add `filters`, `fields`, `sort`, `distinct`, or pagination as needed.

Example flow:

find_datasets({ "query": "מחיר למשתכן" })
get_dataset_info({ "dataset": "mechir-lamishtaken" })
search_records({
  "resource_id": "7c8255d0-49ef-49db-8904-4cf917586031",
  "limit": 5,
  "include_total": true
})

Discover Tags

search_tags({ "keyword": "דיור", "limit": 5 })
find_datasets({ "query": "תחבורה", "tags": "תחבורה ציבורית" })

Use Interactive Discovery

Requires:

MCP_ENABLE_ELICITATION=true

Then an agent may call:

find_datasets({ "query": "תחבורה", "interactive": true })

Compatible clients may show a form asking the user to narrow results.

Use Client-Side Summaries

Requires:

MCP_ENABLE_SAMPLING=true

Then:

summarize_dataset({ "dataset": "mechir-lamishtaken", "language": "he" })

If Sampling is unavailable, the tool returns the dataset metadata and `sampling.used=false`.

Development

npm install

# Type-check
npm run typecheck

# Lint
npm run lint

# Test
npm test

# Build
npm run build

# Refresh local catalog snapshot
npm run catalog:refresh

Run locally:

# stdio
npm run build
npm run start:stdio

# HTTP
npm run build
npm run start:http

Enable optional features locally:

MCP_ENABLE_ELICITATION=true MCP_ENABLE_SAMPLING=true npm run start:http

On PowerShell:

$env:MCP_ENABLE_ELICITATION="true"
$env:MCP_ENABLE_SAMPLING="true"
npm run start:http

Project Structure

src/
  auth/           Authentication providers and Express middleware
  bin/            stdio and HTTP entry points
  cache/          In-memory TTL/LRU cache
  catalog/        Snapshot validation, indexing, fuzzy search, CatalogService
  ckan/           Typed CKAN API client and CKAN response types
  config/         Zod env config, constants, server identity
  core/           Dependency container, MCP server factory, lifecycle
  data/catalog/   Committed catalog snapshot artifact
  formatting/     JSON response builders and guidance text
  observability/  Pino logger
  prompts/        MCP prompt definitions, templates, registration
  resources/      MCP resources, templates, subscriptions
  services/       Domain services for CKAN data access
  tools/          MCP tool definitions and Zod schemas
  transports/     stdio and Streamable HTTP transports
tests/
  fixtures/       Test fixtures
  unit/           Unit tests
scripts/
  refresh-catalog.ts
docs/
  catalog-discovery-layer.md
  MIGRATION.md

Docker

docker build -t data-gov-il-mcp .
docker run -p 3664:3664 data-gov-il-mcp

With optional features:

docker run -p 3664:3664 \
  -e MCP_ENABLE_ELICITATION=true \
  -e MCP_ENABLE_SAMPLING=true \
  data-gov-il-mcp

Quality

The project is expected to pass:

npm run typecheck
npm run lint
npm test
npm run build

Current implementation includes unit coverage for environment parsing, auth providers, CKAN errors, cache, formatting, catalog text normalization/fuzzy/index/search logic, snapshot validation, services, resources, subscriptions, and HTTP Host/Origin guard.

License

MIT