For developers

Lex is MCP-native: you bring the model, Lex brings the evidence. 10 read-only tools over signed indexes, no key, no account, no rate limit on the endpoint.

Connect

Claude Code

claude mcp add --transport http lex https://law.soufien.lu/mcp

Claude Desktop, Cursor, any MCP client

{ "mcpServers": { "lex": { "url": "https://law.soufien.lu/mcp" } } }

Azure AI Foundry Agent Service, remote MCP is native:

{ "type": "mcp", "server_label": "lex", "server_url": "https://law.soufien.lu/mcp", "require_approval": "never" }

No framework at all, it is JSON-RPC over one POST:

curl -X POST https://law.soufien.lu/mcp -H 'Content-Type: application/json' \
  -d '{ "jsonrpc":"2.0", "id":1, "method":"tools/call",
        "params": { "name":"as_of",
          "arguments": { "work":"eu-eurlex:32016r0679",
            "date":"2019-03-15", "mode":"select", "anchors":"art_33" } } }'

The 10 tools

tool	arguments	answers
as_of	work, date, [language], [mode: full\\|outline\\|select], [anchors]	the text in force on a date. outline lists article anchors only; select returns just the anchors you name, use it, codes are large.
article_history	work, anchor	every distinct text one article has had, as validity intervals, plus renumbering events.
timeline	work	all versions of a work with their validity windows.
diff	work, from, to	what changed between two versions.
in_force_on	date, [publisher], [document_type], [limit], [offset]	everything that applied on a given day.
search	query, [publisher], [as_of], [document_type], [language], [limit]	provision-level full-text search; hits are articles, with permalinks.
provenance	lex_id	source URI, retrieval time, record hash, the append-only observation chain.
coverage	[publisher]	what is held, and what is knowably missing.
cited_by	work, [limit]	which articles point AT this law, from the cross-references the publisher writes into its own text. Answers "what depends on this", "who amended it".
changes_in_period	from_date, to_date, [publisher], [document_type], [order], [limit], [offset]	across the corpus: which works gained versions in a window, and how many. The aggregate counterpart of diff and timeline, which cover one work.

Try the joysticks

This calls the same public endpoint your model would. No key, nothing installed , the JSON below is exactly what an MCP client receives.

↑ pick a tool and press "Call it"

It refuses rather than guesses

Every tool returns an envelope with a status. The refusals are part of the contract: no_version_for_date · unknown_work · unknown_anchor · anchor_not_in_version · no_provision_history · text_withheld · outside_observed_window. Build against them: an empty result and a refusal are different things.

Or skip the API, take the data

Every provision of every version, one row each, licence and attribution inline. 3,167 versions carry full text.

DuckDB, one line, no download:

SELECT * FROM read_parquet('https://github.com/SFHAJJI/lex-articles/releases/latest/download/eu-eurlex-provisions.parquet');

Also published as .jsonl.gz. Python examples (standard library only, no pip install) live in examples/ , load provisions, resolve a point in time, verify the hash chain, call this MCP endpoint.

The repositories

repo	what it is	licence
lex	the engine: ingest, derive, index, MCP server, this site	Apache-2.0
lex-articles	derived layer, one Markdown+JSON record per article per version	CC-BY-4.0
lex-corpus-lu-legilux	evidence layer, Legilux's own files, verbatim, plus signed nightly index	CC-BY-4.0
lex-corpus-eu-eurlex	evidence layer, EUR-Lex/Cellar files, verbatim	EU reuse

Why four? Evidence and derivation are kept apart on purpose: the evidence repos are append-only and never rewritten, so a derived record can always be traced back to the exact bytes a state published. The full model →

Everything on this page is exercised by the project's own test and eval suites, and the endpoint is rebuilt nightly. How it was built → · Verify it yourself →