it-eli-mcp

An MCP server for Italian law, in two layers:

• Legislation via Normattiva, the official portal of the Ministero della Giustizia. It fetches acts as Akoma Ntoso with verifiable URN:NIR and ELI identifiers, and returns the text as it stood on any past date (multivigenza). Live and keyless.
• Constitutional case law (Corte Costituzionale): a local, keyless full-text index of every decision since 1956, each with its native ECLI, built from the Court's official open data.
• Supreme Court case law (Corte di Cassazione): live, keyless full-text search over SentenzeWeb, the Court's own free public search engine - civil and criminal decisions, full OCR text, no local index to build.
• Administrative case law (Consiglio di Stato, C.G.A.R.S., TAR): live, keyless search over the Giustizia Amministrativa portal's own public decision search - one backend for the whole administrative jurisdiction, 3.4M+ provvedimenti with native ECLI and structured full text.

Italy has the EU's largest legal profession by headcount - roughly 240,000 lawyers per CCBE figures, ahead of Germany and Spain. No live, keyless MCP connector covered its statute book. This one does, and it also covers the constitutional case law.

Part of the MateMatic eu-legal-mcp production line: the Italian sibling of de-eli-mcp (Germany) and sejm-eli-mcp (Poland), built on the same architecture and citation contract against the Italian source.

No JSON API. Normattiva does not publish a REST API. This connector drives the same session-based flow a browser uses - resolve a URN, read the act page, fetch its Akoma Ntoso export - so every act comes back as structured XML with its ELI intact. See DISCOVERY.md.

Licence. Italian official legal texts are outside copyright (art. 5, l. 633/1941), and Normattiva declares CC-BY-4.0 for its data from 2026-01-01. This connector relays individual acts on request, with attribution and a source_url; it does not bulk-harvest the database (Normattiva's terms restrict that). Every response carries a dataset_note. (A practitioner's read, not formal legal advice.)

Cassazione is NOT Italgiure/ItalgiureWeb. The Court's full-database search (ItalgiureWeb, 35M+ documents) is free only for the judiciary; practitioners pay a subscription. This connector does not use that service. It queries SentenzeWeb (italgiure.giustizia.it/sncass), a separate, free, public search engine the Court itself publishes, under the Italian Open Data License (IODL 2.0), with no authentication

• verified live during development (see DISCOVERY.md). Coverage is narrower than the full Italgiure database but real: 420K+ civil and criminal decisions with full OCR text.

Legislation tools (Normattiva, live)

Constitutional case-law tools (Corte Costituzionale, local index)

The case-law tools read a local SQLite index that is provisioned automatically on the first call - no setup step. On first use the server downloads a small, sha256-verified pre-built cost.sqlite.gz from the GitHub release (built and attached by CI on each release) and decompresses it, or, if none is published, builds the index from the Court's open data (all decisions since 1956); the index is then cached under ~/.matematic and every later call queries it offline. it_case_stats reports provenance and ingested_at so you can see how fresh it is.

To refresh, or to pre-build it ahead of time (e.g. in an offline deployment), run:

italy-eli-mcp-caselaw-ingest

Tuning env vars: IT_ELI_CASELAW_INDEX_URL (override/disable the pre-built download), IT_ELI_CASELAW_INDEX_SHA256 (pin the checksum), IT_ELI_CASELAW_AUTOBUILD=0 (skip the automatic local build). The legislation tools need no index; they fetch live.

Supreme Court case-law tools (Corte di Cassazione, SentenzeWeb, live)

No ingest step - every call queries SentenzeWeb live. Coverage: civil (snciv, ~190K) and criminal (snpen, ~238K) decisions, including the Labour (szdec:L, ~30K) and Tax (szdec:5, ~59K) sub-chambers, full OCR text.

Administrative case-law tools (Consiglio di Stato, C.G.A.R.S., TAR, live)

No ingest step - every call queries the portal's own public decision search live.

Citation verification (anti-hallucination)

What it reports, per citation and overall:

• Existence with a range hint. A cited article that does not exist comes back as not_found with a hint of what does exist: art. 9999 does not exist in this act; the act has 372 articles, numbered art. 1-372.
• Optional content check. A parenthetical description right after a citation (art. 2043 c.c. (risarcimento per fatto illecito)) is compared with the real provision using a character-trigram match. A mismatch is a review signal (content_mismatch), never a hard block - paraphrase is legitimate.
• Hard semantics. When any cited article does not exist, the result status is HALLUCINATION_DETECTED and the tool response carries isError=true, so a calling model cannot mistake it for a passed check. When the text contains no citations at all, the status is NO_CITATIONS_FOUND - explicitly NOT a verification success.
• Structured gaps. Everything the tool could not verify is listed as a typed gap (out_of_corpus, unparseable_citation, act_unresolvable, upstream_unavailable, comma_not_checkable) instead of being hidden in prose. An act Normattiva cannot resolve is a gap, not a hallucination: existence is unknown, not disproven. Cassazione and administrative-court ECLIs are out of the local corpus; verify those with it_cassazione_search / it_ga_search.

it_verify_citations(text="L'art. 2043 c.c. e l'art. 9999 della legge 241/1990 si applicano.")
→ status: "HALLUCINATION_DETECTED" (isError=true)
  citations: [
    {raw: "art. 2043 c.c.", status: "verified", ...},
    {raw: "art. 9999 della legge 241/1990", status: "not_found",
     range_hint: "art. 9999 does not exist in this act; the act has <N> articles, ..."}
  ]
  gaps: []

The parse-verify-report loop is adapted from chrisryugj/korean-law-mcp (MIT) - see THIRD_PARTY.md. The content matcher uses character trigrams (the Korean original uses bigrams, which fit an agglutinative script; trigrams discriminate better for Italian).

Every legislation response carries the citation contract: eli_uri (e.g. eli/id/1990/08/18/090G0294/CONSOLIDATED), urn (e.g. urn:nir:stato:legge:1990-08-07;241), human_readable_citation (e.g. Legge 7 agosto 1990, n. 241), and source_url.

How it identifies an act

• URN:NIR - urn:nir:{authority}:{type}:{date};{number}. A partial URN with the year alone (urn:nir:stato:legge:1990;241) also resolves, so you can cite an act you only know as "L. 241/1990".
• ELI - eli/id/{year}/{month}/{day}/{code}/CONSOLIDATED, read from the act itself. Never invented.

Examples

it_list_codes()
→ [{key: "codice civile", urn: "urn:nir:stato:regio.decreto:1942-03-16;262", ...}, ...]

it_resolve(act_type="d.lgs", year=2001, number=231)
→ {urn: "urn:nir:stato:decreto.legislativo:2001;231",
   human_readable_citation: "Decreto legislativo n. 231/2001", source_url: "..."}

it_get_text(reference="codice civile", article="2043")
→ art. 2043 c.c. (Risarcimento per fatto illecito), with eli_uri + source_url