LexProvenance - lex-provenance-mcp

Verifiable, provenance-first legal corpora for AI systems. A read-only Model Context Protocol connector that lets AI agents search, fetch, and cite neutral, provenance-tracked bodies of law - with a tamper-evident content hash on every citation.

Status: v0.1 - public scaffold. This is a deliberately thin, credible interface layer, not a finished product. See the roadmap and what this is not.

Why this, not a generic legislation feed?

Most legal connectors give an AI the current text of a law and a link. This one gives it a verifiable, point-in-time, regime-curated record - the difference between a citation you can defend and a snippet you have to double-check:

• Tamper-evident - every provision carries a SHA-256 content hash you can re-check, not just a URL.

• Point-in-time - dated versions let you cite the law as it stood on a given day, not only "current".

• Structured & neutral - a strict wall between official text and generated content, plus neutral concept tags and cross-references.

• Authentic-language - the official text in its own language is the record; any translation is labelled unofficial, so the binding text is never ambiguous (BBNJ ships all six authentic UN languages).

• Curated by regime - the assembled, verified corpus for a specific body of law (an AML or digital-asset regime), not a firehose of every Act.

It's also corpus-agnostic: one interface over any provenance-first corpus that exposes a standard manifest - three public "global-commons" corpora today, any regime tomorrow. The connector is the interface; the corpora are the asset.

Related MCP server: Jusratio Case File

What it connects (today)

This repo ships only sample manifests (a few records each) that point to those canonical sources. It does not contain the full corpora.

Tools (read-only)

Every tool is read-only. There are no write, edit, ingest, delete, or shell tools - by design, and enforced by a test (tests/test_read_only.py). See docs/security.md.

How verifiable citations work

get_citation returns the SHA-256 hash and dated version of the exact text it cites - so a citation points not at "Article 14" in the abstract but at a specific, reproducible, tamper-checkable state of Article 14. Change one character and the hash changes, so the citation can be re-verified later. See docs/citation-model.md.

Quickstart

pip install -e ".[dev]"          # pure-Python tool logic + tests, no MCP SDK needed
python -m pytest -q              # all green

# To actually serve over MCP (stdio):
pip install -e ".[mcp]"
lex-provenance-mcp               # or:  python -m lex_provenance_mcp.server

Try the tool logic directly (no SDK required):

from lex_provenance_mcp import tools
print(tools.list_corpora())
print(tools.search_corpus("liability", corpus_id="space_law"))
print(tools.get_citation("space_law", document_id="outer-space-treaty", provision="Article VI"))

Commercial boundary

This connector is the free razor. The enriched, connected, and hosted layers are a separate paid offering - see docs/commercial.md. In short:

• Free / open: the connector, schemas, sample manifests, docs, in-corpus search/fetch/cite.

• Not in this repo (premium): full enriched corpora, cross-corpus reasoning, embeddings, monitoring / treaty-status intelligence, hosted search, private-corpus onboarding.

What this is not

• Not the corpus-construction methodology - that is a separate, private toolkit. This repo reveals what you fetch, never how the record is built.

• Not legal advice. See docs/legal-disclaimer.md.

Roadmap

• v0.1 (this): read-only connector scaffold, schemas, sample manifests, in-corpus tools, verifiable citations, adapter placeholders. No hosted service, no full corpus, no premium layer.

• v0.2: local search over selected public corpora; adapter examples that actually run.

• v0.3: richer citation model + corpus manifests consumed directly from canonical repos.

• Later (premium, not here): hosted MCP endpoint, cross-corpus intelligence, monitoring, institutional access.

Renaming

The public identity lives in one file - mcp_server/identity.py. Change BRAND_NAME, REPO_SLUG, TOOL_NAMESPACE, PACKAGE_NAME there (and rename the folder) to switch to any alternative name; nothing else hard-codes it.

Licence

Code and schemas: Apache-2.0 (LICENSE). Docs: CC BY 4.0 unless stated. Source legal texts are official public documents and are never relicensed. See NOTICE.