eurlex-mcp-server

Search and retrieve EU law via the EUR-Lex Cellar API -- an MCP server that gives AI assistants direct access to EU regulations, directives, court decisions, and more. No API key required.

What You Can Do

Ask your AI assistant questions like:

• "Find EU regulations about artificial intelligence from 2024"

• "Show me the full text of the AI Act (32024R1689)"

• "What EuroVoc topics are assigned to the GDPR?"

• "Which EU acts are about cybersecurity as a topic, even if the word isn't in the title?"

• "Which documents cite the Digital Services Act?"

• "Search for directives about renewable energy"

• "Get the consolidated version of Regulation 2016/679 (GDPR)"

• "Which CJEU judgments interpret the GDPR?"

• "Look up the Court of Justice ruling with ECLI ECLI:EU:C:2014:317"

• "How did Germany transpose the NIS2 Directive?"

• "Outline the AI Act, then show me just Article 5"

• "Give me the plain-language summary of the Digital Services Act"

Related MCP server: Entscheidsuche MCP Server

Features

• 11 specialized tools for searching, fetching, and analyzing EU legal documents (see Tool Reference)

• EuroVoc thesaurus search -- find documents by EU taxonomy concepts, with automatic label fallback across all 24 languages

• CJEU case law -- find judgments, orders, and Advocate General opinions by ECLI, CELEX, party/title, or the act they interpret

• National transposition measures (NIM) -- list how member states transposed a given directive into national law

• Document outline + targeted reading -- get an act's table of contents with plain-text offsets, then jump straight to a specific article with eurlex_fetch; for CJEU case law (CELEX sector 6) the outline also lists each numbered judgment paragraph as Paragraph N, so you can jump to a specific paragraph the same way (paragraph detection is language-independent)

• Plain-language summaries -- retrieve the EU's LEGISSUM summary of an act

• Raw SPARQL escape hatch -- run read-only SELECT/ASK queries directly against Cellar for questions the other tools can't express

• Consolidated versions -- retrieve the latest in-force text of regulations, directives, and decisions, identified by CELEX ID or by doc type + year + number

• Citation graph -- explore which documents cite or are cited by a given act, with a balanced split between the two directions

• Structured metadata -- dates (with null instead of Cellar's 9999-12-31 sentinel for open-ended validity), in-force status, authors, legal basis, EuroVoc descriptors, and directory codes

• Flexible identifiers -- eurlex_fetch, eurlex_metadata, and eurlex_structure accept a CELEX ID, an ELI (e.g. reg/2016/679), or a post-2023 Official Journal reference (e.g. OJ:L_202401689)

• Structured output -- every tool returns a machine-readable structuredContent payload validated against a published outputSchema, alongside the JSON text block

• Offset-based pagination -- eurlex_fetch, eurlex_consolidated, and eurlex_summary return next_offset so long documents can be read in successive calls

• All 24 official EU languages -- request titles and full text in any official EU language (default German)

• No API key required -- uses the public EUR-Lex Cellar SPARQL endpoint

• Resilient by default -- automatic retry with backoff on transient Cellar errors, and in-process caching of EuroVoc labels, consolidated-CELEX lookups, and metadata to cut latency on repeat requests

Quick Start

pnpm dlx eurlex-mcp-server

Or with npx:

npx -y eurlex-mcp-server

Configuration

Claude Desktop

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "eurlex": {
      "command": "npx",
      "args": ["-y", "eurlex-mcp-server"]
    }
  }
}

Claude Code

claude mcp add eurlex-mcp-server -- npx -y eurlex-mcp-server

VS Code

Add to .vscode/mcp.json:

{
  "servers": {
    "eurlex": {
      "command": "npx",
      "args": ["-y", "eurlex-mcp-server"]
    }
  }
}

Cursor

Add to .cursor/mcp.json:

{
  "mcpServers": {
    "eurlex": {
      "command": "npx",
      "args": ["-y", "eurlex-mcp-server"]
    }
  }
}

Windsurf

Add to ~/.windsurf/mcp.json:

{
  "mcpServers": {
    "eurlex": {
      "command": "npx",
      "args": ["-y", "eurlex-mcp-server"]
    }
  }
}

HTTP Transport (Remote Deployments)

When running the server over HTTP (pnpm start:http / dist/http.js) instead of stdio, it's exposed to any client that can reach it over the network. To protect against DNS rebinding attacks, set these environment variables:

Important: If your server runs behind a reverse proxy or load balancer, ensure it forwards the original Host header unmodified (e.g. nginx proxy_set_header Host $host;), otherwise MCP_ALLOWED_HOSTS validation will reject all legitimate traffic — the SDK compares the raw Host header as an exact string.

Production example:

MCP_ALLOWED_HOSTS=mcp.honeyfield.at

This protection is opt-in. If MCP_ALLOWED_HOSTS is not set, the server starts as before and logs a one-line startup warning (MCP_ALLOWED_HOSTS not set — DNS rebinding protection disabled). Any public deployment should set MCP_ALLOWED_HOSTS to its public hostname(s).

Tool Reference

The server exposes 11 read-only tools. Every tool returns both a JSON text block and a machine-readable structuredContent payload validated against a published outputSchema, so MCP clients can consume either representation. Tools that return titles or document text accept a language parameter taking any of the 24 official EU languages as a Cellar 3-letter code (default DEU): BUL, SPA, CES, DAN, DEU, EST, ELL, ENG, FRA, GLE, HRV, ITA, LAV, LIT, HUN, MLT, NLD, POL, POR, RON, SLK, SLV, FIN, SWE.

eurlex_search

Searches EU legal acts by title substring -- a contiguous, case-insensitive phrase match against the document title, not tokenized full-text search. For thematic discovery (the term may not appear in the title) use eurlex_by_eurovoc instead. Results are sorted newest-first within the fetched sample: for very broad single-word queries this is not guaranteed to be the globally newest match -- narrow with resource_type or date_from/date_to if that matters. The response no longer echoes the internal SPARQL query.

eurlex_fetch

Retrieve the full text of a document, identified by exactly one of celex_id, eli, or oj_ref. Long documents are paginated: the response includes returned_chars, total_chars, and next_offset (pass it as the next call's offset to keep reading; next_offset is null once there's nothing left).

* Provide exactly one of celex_id, eli, or oj_ref.

eurlex_metadata

Retrieve structured metadata for a document: document/entry-into-force/end-of-validity/transposition dates, in-force status, authors, legal basis, EuroVoc descriptors, and directory codes. Identified by exactly one of celex_id, eli, or oj_ref (same identifier inputs as eurlex_fetch).

* Provide exactly one of celex_id, eli, or oj_ref.

Notes on the response:

• authors lists the resolved agent names (e.g. "European Parliament", "Council of the European Union") instead of an empty array.

• legal_basis lists the CELEX IDs of the acts this document is based on.

• Date fields (date_document, date_entry_into_force, date_end_of_validity, date_transposition) are null when absent -- including Cellar's 9999-12-31 sentinel for acts with no defined end of validity, which is normalized to null.

• directory_codes are human-readable ("{code-tail}: {label}", where code-tail is the fragment after the last / of the directory-code URI), not raw URIs.

eurlex_citations

Explore the citation graph of a document -- which acts it cites, which acts cite it, and amends/based-on/repeals relations.

With direction: "both", the two directions are queried and split evenly (roughly limit / 2 each) so that a burst of recent cited_by entries can't crowd out cites results. The response includes a counts: { cites, cited_by } object reporting how many of each were actually found.

eurlex_by_eurovoc

Find documents by EuroVoc thesaurus concept (label or URI).

eurlex_consolidated

Retrieve the consolidated (in-force) version of a regulation, directive, or decision. Identify the act with either celex_id or doc_type + year + number -- provide exactly one of the two forms. Like eurlex_fetch, the content is paginated via offset/max_chars/next_offset. The response also includes consolidated_celex (e.g. "02016R0679-20160504") and consolidation_date ("2016-05-04", parsed from that CELEX's date suffix; null if the resolved CELEX has none).

* Exactly one of celex_id or the doc_type+year+number triple must be provided.

eurlex_case_law

Search Court of Justice of the EU case law -- judgments, orders, and Advocate General opinions of the Court of Justice and the General Court. Provide at least one of query, celex_id, ecli, or related_celex (they may be combined). Each result gives celex, ecli, title, date, type, and eurlex_url. Use this for case law; eurlex_search covers only legislation.

* Provide at least one of query, celex_id, ecli, or related_celex.

eurlex_transposition

List the national implementing measures (NIMs) EU member states enacted to transpose a directive into national law -- for transposition and compliance tracking. Each result gives the member state, the national measure's title (in that state's own official language, not translated), its date, the sector-7 NIM CELEX, and a EUR-Lex URL. total_found reports the full count; results is capped at limit. Regulations and decisions generally have no NIMs.

eurlex_structure

Return the outline (table of contents) of an act -- its chapters, sections, articles, and annexes -- each with a character offset into the document's plain text. Use it as a map for targeted reading: read an article's offset, then call eurlex_fetch(celex_id, format:"plain", offset, max_chars) to jump straight to that article. Pass the same language to the follow-up fetch and keep format:"plain" -- offsets are language- and plain-text-specific. Heading recognition covers English, German, and French. For case-law documents (CELEX sector 6, e.g. CJEU judgments) the outline additionally lists each numbered judgment paragraph as "Paragraph N" (level 4); numbered-paragraph detection is language-independent (it keys on the paragraph number, not heading words), so you can get the offset of e.g. paragraph 72 of a judgment and jump eurlex_fetch straight to it. Each outline entry has level (1=part/title/annex, 2=chapter, 3=section, 4=article/paragraph), label (e.g. "Article 5", "Paragraph 72"), title, and offset; total_headings is the full count and the list is capped at 300 for very large acts (truncated=true).

* Provide exactly one of celex_id, eli, or oj_ref.

eurlex_summary

Return the EU's own plain-language "summary of legislation" (LEGISSUM) for an act -- a good quick overview before reading the full legal text. Paginated like eurlex_fetch (offset/max_chars/next_offset). total_summaries is 0 when no summary exists (many acts have none); when several exist, the primary summary's text is returned and the rest are listed in other_summaries.

eurlex_sparql

Expert escape hatch: run a raw, read-only SPARQL query directly against the Cellar endpoint for questions the other tools cannot express. Requires knowledge of the CDM ontology -- see the eurlex_guide prompt for a property cheat sheet. Read-only: only SELECT and ASK are accepted; SPARQL Update and federated SERVICE clauses are rejected. A SELECT with no top-level LIMIT gets LIMIT 50 appended; a top-level LIMIT above 100 is rejected. The response carries vars/bindings (SELECT) or boolean (ASK) plus row_count, returned_rows, and a truncated flag (bindings are dropped whole to stay within a ~40,000-char budget).

CELEX Number Schema

CELEX identifiers uniquely identify EU legal documents. The format is:

[sector][year][type][number]

• Sector (1 digit): 3 = legislation, 6 = case law, 7 = national transposition measures (the sectors these tools cover)

• Year (4 digits): year of the document

• Type (1-2 letters): R = regulation, L = directive, D = decision, J = judgment, etc.

• Number: sequential number

Examples:

Development

Setup

git clone https://github.com/Honeyfield-Org/eurlex-mcp-server.git
cd eurlex-mcp-server
pnpm install
pnpm build

Commands

Testing

pnpm test              # unit tests
pnpm test:integration  # integration tests (hits real API)

Limitations

• Rate limits: The EUR-Lex Cellar API is public but may throttle excessive requests.

• Document availability: Not all documents have full text in all languages.

• Consolidated versions: Only available for regulations, directives, and decisions.

• Response size: Full text is returned per call in max_chars slices (default 20,000 characters) to stay within LLM context limits -- use offset/next_offset on eurlex_fetch/eurlex_consolidated/eurlex_summary to read the rest.

• SPARQL timeouts: Complex queries may occasionally time out on the Cellar endpoint despite the built-in retry with backoff; narrow broad eurlex_search/eurlex_by_eurovoc queries with resource_type or date filters if this happens.

• Search ordering: eurlex_search results are sorted newest-first within the fetched sample only -- for very broad queries this is not guaranteed to be the single globally newest match.

Contributing

See CONTRIBUTING.md for development setup, architecture overview, and submission guidelines.

License

MIT