@cyanheads/eur-lex-mcp-server
Search EU legislation, CJEU case law, and treaties; traverse the CELLAR relationship graph; resolve EuroVoc concepts via MCP. STDIO or Streamable HTTP.
7 Tools • 2 Resources • 1 Prompt
Overview
EU legislation, CJEU case law, and treaties from EUR-Lex. Search documents and case law, fetch full text, resolve citations, trace how acts amend and cite each other, and browse the EuroVoc subject thesaurus.
Tools
Resources
All resource data is also reachable via tools.
Prompts
Skills
The skill is in skills/. Plugin installs, the .mcpb bundle, the npm package, and the Docker image include it, and the server serves it to MCP clients through the Skills extension (skills/list, skill:// resources).
Capability reference
eurlex_search_documents tool
- At least one filter:
keyword (English titles, plus CELEX numbers or fragments such as 2016R0679; a keyword with letters and no digit, such as R(01), scans every CELEX and can take tens of seconds; no body search), document_type (REG, DIR, DEC, TREATY, JUDG, OPIN_AG, PROP, REC), date_from/date_to, eurovoc_concept (from eurlex_browse_subjects), author_institution, or in_force (true/false; false covers repealed, expired, and not-yet-in-force acts)
- Pages of up to 100 via
offset/limit (a larger limit is served at 100 and reported in limit_clamp), newest first; each row flags is_consolidated and is_corrigendum, and corrigenda and consolidated texts appear only under include_corrigenda and include_consolidated. Paging stops at row 10,000: the last page carries no next_offset and a notice giving the date_from/date_to window that continues past it
- No match returns an empty page with a
notice on how to broaden the filters; for eurovoc_concept with document_type JUDG or OPIN_AG, where CELLAR indexes no Court of Justice or General Court ruling by EuroVoc concept, the notice instead points to eurlex_get_cases subject; typed errors: no_filters, invalid_date_range, invalid_author_institution and invalid_keyword (no letters or digits), offset_past_window (offset 10,000 or more)
author_institution matches English authority labels and national-court names returned by eurlex_get_document, such as European Parliament or Supremo Tribunal de Justiça, with or without date bounds
eurlex_get_document tool
- Exactly one of
celex_number, eli_uri, or work_uri (pasted forms such as a CELEX: prefix or EUR-Lex URL are accepted); body as html (default), markdown, or xml (Formex4) in any of the 24 EUR-Lex languages, falling back to English. Case law also carries ecli, formation, referring_court, subject_matter, and case_reference, plus countries, procedure_types, subjects, and procedure_languages where CELLAR records them, as eurlex_get_cases rows carry them; a consolidated text reports its base act as base_act_celex
content_mode "paged" (default), "full", or "metadata_only", capped at 100,000 characters per call (a larger limit is reported in limit_clamp) with content_chars_total/has_more to page on; outline: true lists headings with offsets, and select (e.g. { articles: "1,5,17" }) returns just those sections. For judgments and AG opinions the outline lists section headings, and select: { operative_part: true } returns the ruling aloneis_superseded flags a newer consolidated version (current_consolidated_celex, consolidated_as_of), and resolve: "current_consolidated" serves it; in_force: false comes with repealed_by, end_of_validity, or entry_into_force where CELLAR records one. Typed errors: invalid_identifier_args, not_found, content_challenge (a WAF bot-challenge in place of text)- Citation facts, each absent when CELLAR records none:
eli (a consolidated text keeps its own point-in-time ELI); oj_reference, the Official Journal reference (OJ L 119, 4.5.2016, p. 1–88; no series letter before 1968, a class letter as in OJ L 42I, OJ L, 2024/1689, 12.7.2024 for act-by-act publication), with pages from the English edition or, before 1973, the original-language editions, and the earliest issue for a reprinted act; dates_of_effect, every recorded date as { date, type?, provision? } with type one of entry_into_force, application, partial_application, provisional_application, takes_effect, takes_partial_effect (an act applying in stages lists its general date of application as application and each date covering only some provisions as partial_application) and provision the clause that sets the date, numbered as CELLAR records it (Article 113(a), Article 93.1); and transposition_deadlines for a directive. entry_into_force keeps its meaning of an entry still ahead on a not-in-force act. A consolidated text reports its base act's oj_reference, dates_of_effect, and transposition_deadlines
procedures lists the interinstitutional procedures an act belongs to, reached from a proposal or from the act a procedure adopted, each as { reference, status?, withdrawal_date?, adopted_acts? }: reference in its cited form (2021/0106(COD), which eurlex_lookup_celex resolves), status one of adopted, withdrawn, pending, not_adopted from the flag CELLAR sets (absent when it sets none or several, so a missing adopted_by edge or status never means pending), withdrawal_date on a procedure flagged withdrawn, and adopted_acts the CELEX numbers it produced. A consolidated text reports its base act's; case law belongs to none- Available legacy HTML/Markdown bodies may carry
content_omissions: { tables }: missing-table markers counted across the full source, including outside the returned window. Paging cannot recover these tables. alternative_representations lists PDF streams advertised for the served work and language when discovery succeeds; XML availability remains independent. Metadata-only calls do not inspect source omissions
- National implementing measures carry sparse
national_measure citation fields: journal, number, pages, publication_date, and notification_date, including metadata-only calls. Numbers and pages retain their source spelling; publication, notification, and document dates are distinct, and each is omitted where CELLAR records its 1001-01-01 placeholder for an unknown date. When the body is unavailable, use the gazette citation to retrieve it; no national-portal URL is inferred, and notification does not prove complete transposition
eurlex_lookup_celex tool
- A CELEX number, ELI URI, ECLI, or OJ citation naming its act type and year (
Regulation (EU) 2016/679, Directive 95/46/EC, optionally followed by its parenthesized OJ reference, Directive 95/46/EC (OJ L 281, 23.11.1995, p. 31), or by the period ending a sentence); pasted forms (a CELEX: prefix, a EUR-Lex page URL, an ELI with or without scheme or language segment) are normalized and echoed in celex_number. A citation without its act type (95/46/EC) or year (Regulation No 17) is not parsed
- An Official Journal reference resolves to the act starting on the cited page: issue, publication date, and first page (
OJ L 281, 23.11.1995, p. 31, also of 23.11.1995, a page range, a class letter as in OJ L 42I, OJ C, and before 1968 the series-less OJ 13, 21.2.1962, p. 204/62); the year-first OJ 1995 L 281, p. 31, OJ 1995 L 281/31, and [1995] OJ L281/31; and act-by-act OJ L, 2024/1689, 12.7.2024 and OJ C, C/2024/2664, 29.4.2024. The spellings EUR-Lex act texts and the Court use read the same: OJ No L 281, O.J. L 281, a spaced date (23. 11. 1995), pp. 1–88 or p 31, and OJ L 2024/1689 without its comma. A cited date must equal the issue's publication date. Pages are matched on the English edition, or on any language before 1973 when the issue has no English pages. A reference without its year or first page (OJ L 281, OJ L 119/1) is not parsed. Recognized only under auto
- A Commission document number resolves to the work carrying it:
COM(2021) 206 final, also COM(2021)206, COM/2021/206, final/2 for a later version, a two-digit year as 19YY (COM(90) 314 final), and SEC, SWD, and JOIN numbers (not C(…) decisions). An interinstitutional procedure reference resolves to the act that initiated it: 2021/0106(COD), 2021/106 (COD), or 2021/0106/COD; one without its type is not read. Recognized only under auto
identifier_type auto-detects the format or sets it; ambiguous_identifier fires when auto-detection can't classify the input (including an OJ reference without its year or first page). A case number (C-131/12) returns found: false with a notice pointing to eurlex_get_cases- Returns work URI, confirmed CELEX number, resource type, date, and the case's ECLI;
found: false with a notice for a well-formed identifier that matches no work, for an OJ reference whose date or page matches nothing, for one whose first page starts several acts, and for a document number held in several parts or a procedure with several initiating acts (the notice names each CELEX)
eurlex_get_cases tool
- Filters:
case_number (C-131/12, T-22/20, F-12/05, or a pre-1989 26/62; reaches every judgment, order, and AG opinion filed under it), court (CJEU or GC), case_type (judgment, order, ag_opinion), keyword (English titles, plus CELEX numbers or fragments such as 2013CJ0131), and date_from/date_to; primary records only unless include_derivative adds notices, abstracts, summaries, and corrigenda
- Case-law facts CELLAR records:
member_state (alpha-3 or two-letter code, any case; the referring court's state on a preliminary reference, the defendant on an infringement action, the applicant's state on a direct action), procedure_type (preliminary_ruling, annulment, staff_case, appeal, failure_to_fulfil_obligations, damages, interim_measures, appeal_against_penalty, failure_to_act, arbitration_clause), outcome (successful, unfounded, inadmissible, decision_unnecessary; with procedure_type, on the same procedure), and subject (a phrase matched against the English labels of CELLAR's case-law subject classification, every word required: "trade mark" reaches "Trade marks"; query_echo.subject_labels names the subjects matched)
- Pages of up to 100 via
offset/limit (a larger limit is served at 100 and reported in limit_clamp), newest first; each case carries its ECLI where recorded, countries, procedure_types, subjects, and procedure_languages where CELLAR records them, plus formation, advocate_general, display_title, parties, referring_court, subject_matter, and case_reference, and the raw title only when those fields miss part of it. Paging stops at row 10,000: the last page carries no next_offset and a notice giving the date_from/date_to window that continues past it
- No match returns an empty page with a
notice on how to broaden the filters, as does a subject matching no label (then no case is searched); typed errors: invalid_case_number, invalid_date_range, invalid_keyword and invalid_subject (no letters or digits), offset_past_window (offset 10,000 or more)
case_type: "ag_opinion" includes Advocate General opinions (OPIN_AG) and views (VIEW_AG); resource_type distinguishes AG Opinion from AG View
eurlex_get_relations tool
- Exactly one of
celex_number or work_uri; relation_types narrows to any of cites, amends, amended_by, repeals, repealed_by, implicitly_repeals, implicitly_repealed_by, legal_basis, basis_for, interprets, interpreted_by, consolidated_version, national_transposition, adopts, adopted_by (omit for all fifteen)
basis_for lists the works naming the act among their legal bases; interpreted_by lists the CJEU and General Court rulings interpreting it, and interprets what a ruling interprets; adopts lists the works an act adopts (usually its Commission proposal) and adopted_by the act that adopted a work. member_states narrows national_transposition to named member states (ISO 3166-1 alpha-3 or two-letter codes, echoed as alpha-3); using it without national_transposition fails as member_states_not_applicable- A first page (
offset 0) that traverses national_transposition carries transposition_summary, counted over every measure rather than the page: by_member_state lists each member state with measures (GBR included) with its measure count and newest latest_date, and member_states_without_measures names the states in scope with none — the member_states filter, else the 27 current member states. A missing record is not proof of non-transposition. It is absent when no state in scope has a measure; a count CELLAR fails leaves the page without it, named in the notice
article narrows interpreted_by to rulings with a recorded provision in one article of the act (17, 6a, Article 17, Art. 17, any case; blank is omitted; a paragraph, range, or Roman numeral is rejected). It matches at any depth, ignores leading zeros, and a letter also matches its Latin ordinal (28c reaches Article 28 quater); rulings CELLAR records no provision for are left out, as the notice says, and the echoed article is the normalized form. Using it without interpreted_by fails as article_not_applicable- Amendment, repeal, interpretation, and adoption rows carry
provisions, the provisions CELLAR records the edge as targeting, in compact codes (A17P1LC is Article 17(1)(c)) or authority-token codes (AR 43 PA 4); a row without them records none, which does not mean the edge reaches the whole act. A provisions lookup CELLAR fails returns the rows without them, named in the notice
- One hop, paged per relation type and direction via
offset/limit (up to 100, default 10; a larger limit is reported in limit_clamp), newest first. limit is the only page boundary: every type and direction returns up to that many whole rows, titles and provisions included, and next_offset is the offset plus the limit served. Each type and direction pages to row 10,000, then a notice routes older works to eurlex_query_sparql with a ready-to-run query. Omitting relation_types equals listing all fifteen types
- Each relation carries
relation_type, direction (outgoing/incoming), related_work_uri, and where known related_celex_number, related_date, related_title (English only), and related_member_state on national_transposition rows; empty_relation_types separates "no edges of this type" from "paged out". Typed errors: invalid_identifier_args, member_states_not_applicable, article_not_applicable, not_found, offset_past_window
eurlex_browse_subjects tool
- Matches preferred and alternative EuroVoc labels, so a common synonym resolves to its concept, in any EU official
language (default English); offset/limit pagination (up to 50; a larger limit is served at 50 and reported in limit_clamp), to row 10,000 (offset_past_window beyond)
- Returns concept URI, preferred label, code, broader (parent) label, and the alternative label that matched when one did; an empty page with a
notice when nothing matches
- Exact label matches rank first, then label or word-start matches, then other substring matches, so
"AI" leads with artificial intelligence
eurlex_query_sparql tool
- Read-only SELECT only (update forms and ASK/CONSTRUCT/DESCRIBE are rejected);
cdm:, skos:, and xsd: prefixes are auto-injected; optional timeout_hint (1000–55000 ms)
- Results capped at
MAX_SPARQL_RESULTS (default 100); a truncated result's notice says to add ORDER BY and repeat with a higher OFFSET, up to row 10,000. A zero-row result from an untyped string literal gets a notice to type it (^^xsd:string, ^^xsd:anyURI for an ELI) or language-tag it
- Typed errors:
not_read_only (a SPARQL Update), unsupported_query_form (ASK, CONSTRUCT, DESCRIBE, or no SELECT), sparql_error, sparql_timeout
eurlex://document/{celexNumber} resource
- Metadata snapshot as
application/json — resource type, author institution(s) labelled as eurlex_get_document labels them, Advocates General, date, English title, in-force flag, eli, oj_reference, dates_of_effect, transposition_deadlines, legal basis, EuroVoc subjects, and procedures (the four citation fields and the procedures as in eurlex_get_document); a consolidated text keeps its own eli, adds base_act_celex, and reads authors, in-force flag, oj_reference, dates_of_effect, transposition_deadlines, legal basis, subjects, and procedures from that act
celexNumber comes from eurlex_search_documents, eurlex_get_cases, or eurlex_lookup_celex
eurlex://document/{celexNumber}/relations resource
- One-hop relationship summary over all fifteen
eurlex_get_relations types, with no provisions, each related work with its related_date and English related_title where it has them, capped at 25 per relation type and direction (newest kept)
truncated plus a continuation pointer to eurlex_get_relations when more relations exist, its input naming the CELEX and offset 25 so the tool continues each type and direction where the summary stopped
eurlex_comparative_analysis prompt
- Arguments:
domain required; focus optional, folded into its matching analysis axis or added as its own section
- Returns a research plan that restates the domain in EU legal vocabulary ("antitrust" → "competition law"), chains
eurlex_browse_subjects → eurlex_search_documents → eurlex_get_document → eurlex_get_relations for the EU side and courtlistener_search_opinions for the US side, plus a six-axis analysis framework
Features
Built on @cyanheads/mcp-ts-core: stdio and Streamable HTTP transports, pluggable auth (none / jwt / oauth), swappable storage (in-memory, filesystem, Supabase, Cloudflare KV/R2/D1), structured logging with optional OpenTelemetry tracing.
EUR-Lex-specific:
- No API key required — CELLAR SPARQL and the EUR-Lex REST content endpoints are both publicly accessible
- SPARQL is POSTed with CDM prefix declarations built in; server-side LIMIT enforcement (max 100) guards against Virtuoso timeouts
- Act text is fetched via CELLAR content negotiation; HTML passes through as served, multi-part Formex4 XML is assembled into one document, and Markdown is converted server-side
- Syntax, compile, and sort-window errors fail at once as
sparql_error (a sort-window error, SR353, hints to lower OFFSET or narrow the query) and timeouts as sparql_timeout, neither retried within the call; a sparql_timeout for running out of time is marked retryable, since a later call can finish, and Virtuoso's refusal of a too-costly query is not. Gateway pages and other transient failures are retried inside the call's shared query budget (SPARQL_QUERY_TIMEOUT_MS), and a retry ladder the budget cuts short reports the last attempt's own failure
- Automatic English fallback when a requested translation is unavailable, with requested/effective language reported
Agent-friendly output:
- EuroVoc prerequisite guidance in server-level instructions — agents are directed to
eurlex_browse_subjects before concept-filtered searches
eurlex_lookup_celex confirms CELEX/ELI/ECLI existence upfront, preventing downstream errors in document or relation fetchescontent_status, content_unavailability_reason, and requested/effective language fields distinguish skipped, available, absent, upstream-failed, and incomplete content without string parsing- Typed
reason codes on every tool's error contract let agents branch on outcomes programmatically
Getting started
Connect to the hosted instance, no install:
Or run it locally:
For other clients, add this to the MCP config file:
No API key is required. For Docker or a source build, see Running the server.
Configuration
All configuration is validated at startup via Zod schemas in src/config/server-config.ts.
See .env.example for the full list of optional overrides.
Running the server
Local development
Requires Bun v1.4.0 or higher (or Node.js v24+).
-
Clone and install:
-
Build and run:
-
Run checks and tests:
Docker
The image defaults to HTTP transport, stateless session mode, and logs to /var/log/eur-lex-mcp-server. To run it over stdio from an MCP client, use "command": "docker" with "args": ["run", "-i", "--rm", "-e", "MCP_TRANSPORT_TYPE=stdio", "ghcr.io/cyanheads/eur-lex-mcp-server:latest"]. To build locally, docker build -t eur-lex-mcp-server . then docker run --rm -p 3010:3010 eur-lex-mcp-server; OpenTelemetry peer dependencies are installed by default — build with --build-arg OTEL_ENABLED=false to omit them.
Project structure
Development guide
See CLAUDE.md for development guidelines and architectural rules. The short version:
- Handlers throw, framework catches — no
try/catch in tool logic
- Use
ctx.log for request-scoped logging, ctx.state for tenant-scoped storage
- Register new tools and resources via the barrels in
src/mcp-server/*/definitions/index.ts
- Wrap external API calls: validate raw → normalize to domain type → return output schema; never fabricate missing fields
Contributing
Issues are welcome. Run checks and tests before submitting:
License
Apache-2.0 — see LICENSE for details.
