README.md

metadata

title: CanLex MCP
sdk: docker
app_port: 7860
pinned: false

CanLex MCP

A remote Model Context Protocol server for Canadian legal research, built for front-line border-services work. The legislation corpus is bilingual: every Justice Laws Act and regulation (84) also carries its parallel French text (the Constitution Acts and the STCA texts are English only), and a French-language query searches both languages automatically. Fourteen read-only tools over streamable HTTP (or local stdio):

• canlex_search_legislation — hybrid (BM25 + semantic) search with a cross-encoder reranker over the whole corpus: federal border, customs, criminal, firearms, drug, food/health, labour and privacy legislation with their Schedules (84 Justice Laws instruments, from IRPA and the Customs Act to the sanctions Acts, the Criminal Records Act, the Income Tax Act, the aviation-security instruments and the public-service labour statutes); the ENF operational manuals, the other active IRCC manuals (CP 14, IN 1-2, IP 3/4/7, OP 8/9/14/22; IP 10 and OP 21 are held back by the eval gate), the 17 IRCC operational bulletins still in force (of 336 IRCC lists) and IRCC program delivery instructions; CBSA D-Memoranda and the AMPS Master Penalty Document; the Charter; the Safe Third Country Agreement; Treasury Board collective agreements; NJC directives; the CBSA Code of Conduct and the Treasury Board people-management instruments; CBSA detention policy and the IRB Chairperson's Guidelines and the IRB Legal Services papers on criminal equivalency and criminal refusals; IRCC's rehabilitation application guide; ~265 leading decisions; delegation instruments; benefit-plan booklets; and CanLex's curated commentary.
• canlex_get_section — fetch one provision verbatim by Act and section (accepts 's. 34', '34(1)(c)', 'Schedule I').
• canlex_us_disposition — step 1 of the IRPA s. 36 screening chain: is a foreign disposition a "conviction"? Curated for US records, with a 51-jurisdiction state-by-state dataset; a non-US record (the optional country, or a state value that is not a US state or territory) gets the framework and the entry-level verdict, labelled "non-US record: curated coverage is thin". The tool names keep their us_ prefix so saved client configs keep working. The answer ends with a READ NEXT block: the official guidance and statute on the point, from the corpus (chunk id, status and a verbatim quote; unofficial ATIP copies of withdrawn chapters flagged and listed after the official sources).
• canlex_us_equivalency — step 2: the Canadian equivalent of a foreign offence (US-curated pairings), its verified maximum penalty, and the inadmissibility branch, with the pairing's READ NEXT block.
• canlex_rehabilitation — step 3: deemed rehabilitation / eligibility to apply, under IRPA s. 36(3)(c) and IRPR ss. 17–18.1 — from dates, with the earliest eligibility dates computed.
• canlex_screen_record — the three steps in one call for a foreign record (US state rows where they exist; optional country), from the record's own words and dates: a screening memo with the draft allegation paragraph (both paragraphs when the disposition or a tie is unsettled), the next steps at the port of entry (s. 44 report, the removal path by paragraph, TRP, rehabilitation application with the IRPR fees, family members by application type), what the chain assumed, an "s. 37 also to consider" block where the offence or the optional organization fact raises organized criminality, a "Read next" line per step, and one structured result line. Where the record's words reach curated content still held for review that the deployment lacks, the memo says so ("HELD FOR REVIEW -- not in this deployment") and lists it among what the chain assumed. A conviction in Canada (country 'Canada' or a province) is screened on s. 36(1)(a)/(2)(a) instead: discharges, conditional sentences, s. 18.1, the record suspension as the cure, the removal path and a permanent resident's appeal line. An optional date_of_birth (or age_at_offence) prints the youth block, statute and IRCC policy kept apart. At the top, THE ANSWER CHANGES WITH THESE FACTS lists each missing fact that changes the branch, the allegation paragraph or the rehabilitation status (the chain re-run with every plausible value; CANLEX_RESULT decisive_missing). An optional conviction_date bounds the maximum at commission; sentence_imposed with sentence_date adds a completion date IF SERVED EXACTLY AS IMPOSED (Guide 5312's rules; never a verdict, and a fine blocks it); detail: "card" returns a result card of about ten lines instead of the memo. The full memo ends with a DRAFT s. 44(1) narrative skeleton in the ENF 5 s. 13.1 form, in English, French or both (language): the proposed IRPA paragraph with its words, the offence as recorded, the Canadian provision (the pairing's curated report provision, chosen by the matched limb and the maximum at commission) and the maximum in the statute's own words. Every bracket is to be verified; what CanLex does not hold (name, place, the conviction date unless given, the status evidence) stays [CONFIRM: ...], and a result that is not settled gets one allegation per reading, never a settled one.
• canlex_screen_history — a whole record at once: every entry through the chain (foreign or Canadian), then the record-level count for IRPR s. 18(2) (equal or missing offence dates on a two-offence count give both readings), deemed rehabilitation tested limb by limb (IRPR s. 18(2)(a)-(c) and s. 18.1: a table with the entry that decides each limb and the date it starts or stops holding; both readings where an unsettled reading decides it; an optional conviction_date per entry, which also bounds the maximum at commission) and the s. 17 clock from the most recent conviction, Canadian ones included. The decisive facts, detail: "card" and the narrative skeleton (one allegation per entry that grounds one; the two-offence limb over the summary-only convictions) as for canlex_screen_record; an entry's relief (a pardon, expungement, sealing or record suspension, or 'none') is read as the disposition's later history.
• canlex_screen — the same record pipeline, with the entries as typed objects (one per conviction, disposition or charge, in the record's own words; an unknown key is refused, so a misspelt key never silently changes an answer). There is no free-text rap-sheet parsing. A record-level relief applies to every entry without its own; 'none' confirms there is none. Returns the result card by default (detail: "full" for everything, the narrative skeleton included). Every screening tool's CANLEX_RESULT carries confidence: what each step rests on (the entry's and pairing's authority status, whether the state row was verified, how many cited decisions are in the corpus, where the maximum came from, the s. 18(2) limbs left unknown).
• canlex_transborder — an offence committed on entering Canada (an undeclared firearm, cannabis or drugs): transborder criminality under IRPA s. 36(2.1) and IRPR s. 19 — whether the offence is prescribed, whether the Minister's delegate or the Immigration Division makes the deportation order (IRPR ss. 228(1)(a.01), 229(1)(d.1)), no rehabilitation, the contraventions check, TRP and a draft allegation. Foreign nationals only.
• canlex_list_acts — the corpus inventory and its currency.
• canlex_whats_new — provisions amended or newly in force since a date.
• canlex_citing — reverse cross-references: what cites a given section.
• canlex_toc — an Act's table of contents (Parts and marginal notes).
• canlex_case — live CanLII case citator (citation graph by neutral citation or URL).

The server also publishes an MCP prompt (us_record_screening, the guided three-step screening flow) and resources (the answering instructions, the corpus inventory and its per-type views, and the conviction, equivalency and rehabilitation methodology texts) that any MCP client can use.

The labour side now carries the full union stack alongside the FB collective agreement: the CIU National By-Laws (EN/FR), CIU Regulations, Branch Rules and local modifications, the Branch Elections Kit, the PSAC Constitution, the union's member guides, and every published site VSSA — all bannered as bargaining-agent documents, distinct from employer policy.

Retrieval runs fully locally inside the container and needs no API key. The optional case citator reads a CanLII key from the CANLII_API_KEY environment variable. A weekly staleness check (canlex.refresh --check, run every Monday by a Claude Code scheduled task on the maintainer's machine -- RUNBOOK.md, "The staleness watch") watches upstream sources for drift; see RUNBOOK.md for the maintainer's refresh choreography.

Add this Space's /mcp URL (https://beemer0-canlex.hf.space/mcp) as a custom connector in any Claude client, or point any other MCP-capable agent at it. CanLex is an MCP server only: it has no chat page of its own, and it calls no hosted model API.

Call log (2026-09-16): tool calls to this server, including their arguments, are logged by the operator, privately, to improve CanLex (RUNBOOK.md, "Call log"). The server says the same to every MCP client when it connects (its MCP instructions) and at its root URL, https://beemer0-canlex.hf.space/.

2026-09-16: the companion Gradio chat page (the separate Canlex-web Space) and its Google Gemini API integration were retired. The webapp/ directory, its tests and the predeploy parity step were removed; the git history keeps them.