io.github.klein-business/legal-text-mcp-de

legal-text-mcp-de

Cite-grade German legal-text infrastructure for LLM agents. ~8 500 federal + Länder + EU laws via MCP, HTTP API, and a typer shell CLI — all over the same runtime, with cryptographic provenance from gesetze-im-internet.de and EUR-Lex / Cellar.

It is local or server-side infrastructure: no SaaS, no billing, no accounts, no tenant model, and no legal advice. The runtime loads either the committed fixture packages used by fast CI or a generated production corpus package built outside Git. Official text comes from gesetze-im-internet.de for German federal laws and from EUR-Lex / Cellar for EU acts such as the GDPR.

No legal advice. This software returns text and structured metadata. It does not interpret the law, advise on it, or produce any legal conclusion. The maintainer assumes no liability for use in legal decision-making contexts.

Older internal documentation has been archived under docs-legacy/summary.md.

Status

Features

• MCP tools for listing laws, fetching norms, resolving citations, full-text search, and source provenance.
• HTTP API (FastAPI) over the same runtime, with structured /health, /ready, /laws, /search, and OpenAPI endpoints.
• Provenance-first design: every law and norm carries source URL, fetch timestamp, content hash, and the parser path it traversed.
• Two corpus modes: committed fixture packages for deterministic tests and CI, or a generated production package built from official sources at runtime.
• No editorial bundling: this repository ships tooling, not legal text. Texts are loaded from official sources at runtime.

Installation

pip or uvx with an explicit dataset

pip install legal-text-mcp-de==2.1.3
DATASET_PATH=/path/to/dataset legal-text-mcp-de serve
# Or:
DATASET_PATH=/path/to/dataset uvx legal-text-mcp-de==2.1.3 serve

• The package and standard Docker image contain the runtime only.
• DATASET_PATH must point to a normalized dataset directory with laws.json, norms.json, and search-index.json.
• Public corpus OCI references and the -full image currently return DENIED; there is no verified public corpus download.
• The CLI and HTTP runtime do not automatically prepare a usable dataset. A .tar.zst bundle is not a supported direct runtime path.
• Build the nine registered federal laws from official sources using the federal corpus guide. This requires the current source checkout and includes a fix beyond v2.1.3.
• prepare_data.build_corpus --sources bund is not implemented. Its archive output is separate from the serving dataset format.

Docker with a mounted dataset

docker run -p 8001:8001 \
  -v /absolute/path/to/dataset:/data/legal-texts:ro \
  ghcr.io/klein-business/legal-text-mcp-de:2.1.3 \
  uv run --frozen --no-sync legal-text-mcp-de http

• Use a dataset compatible with the selected runtime version.
• /health checks the process; /ready checks whether the HTTP API can serve its dataset.
• For freshly built federal data, follow the guide's source-image build instructions.

Mode 5 — Public-hosted service

// claude_desktop_config.json
{
  "mcpServers": {
    "legal-de": {
      "url": "https://mcp.klein.business/legal/de",
      "transport": "streamable-http"
    }
  }
}

CLI

legal-text-mcp-de ships a full subcommand CLI. Bare invocation prints --help. Common commands:

legal-text-mcp-de serve              # start the MCP server (replaces the v2.0 bare invocation)
legal-text-mcp-de http               # start the FastAPI HTTP API
legal-text-mcp-de laws --query DSGVO # list laws
legal-text-mcp-de norm BGB "§ 433"   # fetch a single norm
legal-text-mcp-de search Werbung     # full-text search
legal-text-mcp-de corpus info        # inspect locally cached bundle metadata
legal-text-mcp-de version            # version + Python + platform

Place --json or --output yaml before the subcommand for machine-readable output; the data field contains the HTTP response shape. See CLI reference for the full subcommand list.

BREAKING in v2.1.0: bare legal-text-mcp-de now prints --help. Append serve to keep the v2.0 behaviour (started the MCP server).

Quickstart

MCP-client compatibility

• Tested requires a recorded client version, transport, successful tools/list, and one successful tool call.
• Configuration examples alone are not test evidence. Client runs below remain [UNVERIFIED]; protocol-level CI does not establish desktop-client compatibility.

Run the MCP server with the committed fixture corpus

uv sync --all-groups

DATASET_PATH=tests/fixtures/normalized \
STRICT_STARTUP=true \
uv run legal-text-mcp-de serve

The default transport is streamable HTTP at http://localhost:8001/mcp. The CLI does not expose stdio; see stdio transport status.

Dev shortcut. A Justfile wraps the common uv invocations (just install, just test, just lint, just run, just api, just docs). Install via brew install just. The Justfile is optional — CI only uses uv directly.

Run the HTTP API

Since v2.1 the easiest way is the new CLI:

DATASET_PATH=tests/fixtures/normalized \
STRICT_STARTUP=true \
uv run legal-text-mcp-de http

The default port is 8001 (from settings.port / PORT env var); use --port <N> to override.

Or directly via uvicorn (equivalent — same FastAPI app):

DATASET_PATH=tests/fixtures/normalized \
STRICT_STARTUP=true \
uv run uvicorn legal_text_mcp_de.http_api:app --host 127.0.0.1 --port 8001

The HTTP layer enforces MAX_REQUEST_BODY_BYTES (default 1048576 = 1 MiB) as a defence-in-depth body-size cap; over-limit requests return 413 Payload Too Large. The operator's reverse proxy is still expected to enforce a proxy-level limit on top of this.

Docker

The Docker image does not bundle legal text data. Mount a validated package at /data/legal-texts:

docker run --rm -p 8001:8001 \
  -v /path/to/legal-text-package:/data/legal-texts:ro \
  ghcr.io/klein-business/legal-text-mcp-de:2.1.3 serve

• See examples/docker-compose/http/ for a minimal HTTP-mode Compose example.
• See examples/docker-compose/production/ for a committed, CI-tested production stack (MCP + HTTP behind Caddy, switchable via Compose profiles).

MCP Resources

Since v2.0, the corpus is exposed as read-only legal:// URIs that any MCP client can load directly into its LLM context. Examples:

• legal://laws — paginated index
• legal://laws/bgb — BGB header + norm index
• legal://laws/bgb/norms/par:433 — § 433 BGB as Markdown
• legal://corpus/coverage — what's in the corpus

See docs/concepts/mcp-native.md for the full URI catalogue.

MCP Prompts (slash-commands)

Five curated workflows appear as slash-commands in MCP clients:

Smart Tools

research_topic is a multi-step research tool that orchestrates 2 LLM-sampling calls per invocation:

1. Corpus search for candidate norms
2. LLM ranking of candidates by relevance
3. Related-norms graph loading
4. LLM synthesis of a structured research report

When the client lacks sampling support, the tool returns a degraded report with ranked candidates only.

Data Sources

No text from these sources is committed to this repository. The generated-corpus pipeline fetches them at build time and stores provenance in a manifest.

MCP Tools

See the MCP tools reference for the full surface. Highlights:

• list_laws(query?) — list loaded laws with optional metadata filter.
• get_law(code) — law metadata + normalised norm summaries.
• get_norm(code, norm) — return one structured norm.
• search_laws(query, codes?) — search normalised texts.
• resolve_citation(...) — resolve structured citations without legal interpretation.
• get_source_metadata(code?), get_source_limitations(...), get_corpus_coverage(), get_related_norms(code, norm).

MCP tools return JSON-compatible objects. They do not return double-serialised JSON strings.

HTTP API

Article-plus-section paths must be URL-encoded:

/laws/egbgb/norms/art%3A246a%2Fpar%3A1

Documentation

Full documentation is published at klein-business.github.io/legal-text-mcp-de.

Quick links:

• Quickstart
• MCP tools
• HTTP API
• Operations: security, SBOM, cosign-verify, versioning, threat model
• Roadmap

Source-of-truth documents live in the repo: README.md, CHANGELOG.md, SECURITY.md, CONTRIBUTING.md, GOVERNANCE.md, NOTICE, LICENSE.

Development

uv sync --all-groups
uv run --group dev pytest

The full fixture-backed release gate:

uv run --group dev python scripts/verify_release.py

The public-flip readiness gate:

uv run --group dev python scripts/verify_pre_flip.py

A Justfile wraps the common targets (just test, just lint, just docs, just run, just api) for convenience.

Contributing

See CONTRIBUTING.md for guidelines, code of conduct, and security policy. All contributions must comply with the Developer Certificate of Origin (sign-off with git commit -s).

Verification

Every release since v1.0.0 is signed and accompanied by an SBOM and SLSA-3 provenance. Examples below use v2.1.3; substitute the tag you actually pulled.

Cosign image signature

cosign verify ghcr.io/klein-business/legal-text-mcp-de:2.1.3 \
  --certificate-identity-regexp 'https://github.com/klein-business/.*' \
  --certificate-oidc-issuer https://token.actions.githubusercontent.com

See Verify with cosign for SBOM and SLSA attestation verification.

Licence and acknowledgements

This project is licensed under the Apache License 2.0. See NOTICE for required attribution.

Derived from floleuerer/deutsche-gesetze-mcp (Copyright (c) 2025 Florian Leuerer, MIT). Upstream licence terms are preserved in licenses/MIT-floleuerer.txt.

MCP Registry: io.github.klein-business/legal-text-mcp-de · registry.modelcontextprotocol.io

mcp-name: io.github.klein-business/legal-text-mcp-de