Law Scrapper MCP

A comprehensive Model Context Protocol (MCP) server for accessing and analyzing Polish legal acts from the Sejm API, enabling AI-powered legal research and document analysis.

Features

• Comprehensive legal act access - Full access to Polish legal acts from Dziennik Ustaw (DU) and Monitor Polski (MP)

• Advanced search and filtering - Multi-criteria search by date, type, keywords, publisher, and status

• Result Store with chained filtering - Store search results and filter with regex, type/status/year match, date ranges, sorting

• Document Store pattern - Load acts into memory for efficient section-level navigation and search

• Detailed document analysis - Metadata, structure, references, and content retrieval

• Content processing - Automatic PDF-to-text and HTML-to-Markdown conversion

• Date calculations - Specialized date utilities for legal document analysis

• System metadata - Keywords, statuses, document types, and institution data

• Official MCP SDK - Built with mcp[cli]==2.2.0 and MCPServer; STDIO and stateless Streamable HTTP transports

• Async HTTP client - Efficient httpx client with retry logic and connection pooling

• TTL caching - Intelligent response caching with configurable TTL

• Structured logging - JSON and text log formats for easy debugging

• Docker support - Containerized deployment with docker-compose

• Comprehensive documentation - Examples and clear parameter descriptions

Related MCP server: Polish Law MCP Server

Requirements

• Python: 3.13 or higher

• Package manager: uv (recommended) or pip

• Internet connection: Required for accessing Sejm API endpoints

• MCP-compatible tool: Cursor IDE, Claude Code, or other MCP clients

Installation

Using uv (recommended)

# Clone the repository
git clone https://github.com/numikel/law-scrapper-mcp.git
cd law-scrapper-mcp

# Install dependencies
uv sync

# Install with dev dependencies
uv sync --extra dev

Using pip

# Clone the repository
git clone https://github.com/numikel/law-scrapper-mcp.git
cd law-scrapper-mcp

# Install dependencies
pip install -e .

From a GitHub Release (core and source modules)

Each release attaches the core and every source module as separate wheels, plus checksums.txt. A module requires the core of exactly its own version, so install both from the same release (replace X.Y.Z with the release version):

RELEASE=https://github.com/numikel/law-scrapper-mcp/releases/download/vX.Y.Z
pip install "$RELEASE/law_scrapper_mcp-X.Y.Z-py3-none-any.whl" \
            "$RELEASE/law_scrapper_mcp_voivodeship_journals-X.Y.Z-py3-none-any.whl"
# or: uv pip install with the same two URLs

An installed module is active; see Source modules to switch it off.

Using uvx (no installation required)

For quick testing without cloning the repository:

# Run the server directly from GitHub
uvx --from git+https://github.com/numikel/law-scrapper-mcp law-scrapper

Quick start

STDIO transport (default)

STDIO is the default transport for MCP communication. Start the server and connect from your MCP client:

# Run the server
uv run python -m law_scrapper_mcp

# Or use the installed script
law-scrapper

Configure in your MCP client (e.g., Cursor .cursor/mcp.json):

{
  "mcpServers": {
    "law-scrapper-mcp": {
      "command": "law-scrapper"
    }
  }
}

For Claude Code:

claude mcp add law-scrapper "uvx --from git+https://github.com/numikel/law-scrapper-mcp law-scrapper"

HTTP transport (streamable-http)

Run the server on HTTP with streamable-http transport:

# Run with HTTP transport on port 7683
LAW_MCP_TRANSPORT=streamable-http uv run python -m law_scrapper_mcp

# Or specify a custom port (the host stays on loopback unless you configure
# an authentication mode — see "Authenticated remote deployment" below)
LAW_MCP_TRANSPORT=streamable-http LAW_MCP_PORT=8080 uv run python -m law_scrapper_mcp

Configure in your MCP client:

{
  "mcpServers": {
    "law-scrapper-mcp": {
      "url": "http://localhost:7683/mcp",
      "transport": "streamable-http"
    }
  }
}

Note: The URL must include the /mcp path. The server exposes stateless Streamable HTTP at /mcp, not at the root. Using http://localhost:7683 without /mcp results in 404 (Not Found).

Docker

Build and run with Docker:

# Build the image
docker build -t law-scrapper-mcp .

# Run with STDIO transport (default)
docker run -it law-scrapper-mcp

# Run with HTTP transport on port 7683
docker run -it -p 7683:7683 -e LAW_MCP_TRANSPORT=streamable-http law-scrapper-mcp

Or use docker-compose:

# Run with STDIO transport
docker compose up

# Run with HTTP transport (set TRANSPORT=streamable-http in docker-compose.yml)
docker compose -f docker-compose.yml up

CI builds this image and smoke-tests both transports (/health over streamable-http, one initialize over STDIO) on every code change, so a Dockerfile that no longer builds fails the pull request rather than a downstream registry.

The image ships every released source module installed and switched off (LAW_MCP_SOURCE_MODULES=-*). Enable one with -*,<name>. Never use a bare <name>: it replaces -* and turns every other installed module on as well.

docker run -it -p 7683:7683 \
  -e LAW_MCP_HOST=0.0.0.0 \
  -e LAW_MCP_AUTH_MODE=bearer \
  -e LAW_MCP_AUTH_TOKEN="$(openssl rand -base64 32)" \
  -e 'LAW_MCP_SOURCE_MODULES=-*,voivodeship-journals' \
  law-scrapper-mcp

Configuration

All settings are configured via environment variables with the LAW_MCP_ prefix:

The list-valued settings — LAW_MCP_ALLOWED_HOSTS, LAW_MCP_ALLOWED_ORIGINS, LAW_MCP_TRUSTED_PROXIES, LAW_MCP_AUTH_REQUIRED_SCOPES and LAW_MCP_AUTH_ALGORITHMS — accept either a comma-separated value (a:*, b:*, the form used throughout this document) or a JSON array (["a:*", "b:*"]). Both are equivalent; surrounding whitespace is trimmed.

Logs go to stderr (stdout carries the MCP protocol on the STDIO transport). Every record carries a request_id correlating it with a single tool call — lifespan for records emitted outside one:

2026-08-24 12:00:00,123 - law_scrapper_mcp.tools.search - [a1b2c3d4] - INFO - Stored result set rs_1: 20 results (total 137)

{"timestamp": "2026-08-24T10:00:00.123456+00:00", "level": "INFO", "logger": "law_scrapper_mcp.tools.search", "request_id": "a1b2c3d4", "message": "Stored result set rs_1: 20 results (total 137)"}

Set LAW_MCP_LOG_LEVEL=DEBUG to recover the detail deliberately kept off INFO: search query text and the exception messages behind validation and upstream failures.

Example environment configuration:

export LAW_MCP_TRANSPORT=streamable-http
export LAW_MCP_PORT=7683
export LAW_MCP_LOG_LEVEL=DEBUG
export LAW_MCP_CACHE_METADATA_TTL=86400

Identity towards upstream hosts

Every request the server sends to an upstream host (the Sejm API and any host a source module talks to) carries the same identity header, so that a host's administrator can tell what is calling and reach the person responsible instead of blocking the address:

User-Agent: law-scrapper-mcp/<version> (+https://ops.example.org/contact)
User-Agent: law-scrapper-mcp/<version> (+mailto:ops@example.org)

LAW_MCP_CONTACT changes only the contact; the server name and version are the ones the server reports. The contact is yours to set, either as an http(s):// URL or as an e-mail address (the address is written into the header as mailto:<address>). A source module adds its own module/version token after the contact, so a host also sees which module generates its traffic.

Unset, the contact is this project's repository. That is fine for your own machine, but behind a shared or public deployment it sends complaints about your traffic to the project's authors. When the server listens beyond the loopback over HTTP and still sends the default contact, it logs one warning at startup asking you to set your own. A value that is not a plain URL or e-mail stops startup with a message naming the variable; the rejected value is not echoed.

export LAW_MCP_CONTACT=ops@example.org

Authenticated remote deployment

The HTTP transport binds 127.0.0.1 by default and refuses to start on any other address unless an authentication mode is configured. Two modes exist and neither falls back to the other.

Bearer token — local and simple deployments:

export LAW_MCP_AUTH_TOKEN=$(openssl rand -base64 32)   # min. 32 bytes
LAW_MCP_TRANSPORT=streamable-http \
LAW_MCP_HOST=0.0.0.0 \
LAW_MCP_AUTH_MODE=bearer \
  law-scrapper

In production prefer LAW_MCP_AUTH_TOKEN_FILE=/run/secrets/law_mcp_token — an environment variable is visible in docker inspect and /proc/<pid>/environ. Setting both sources is a startup error, not a precedence rule.

OAuth 2.1 / OIDC — corporate deployments. Works with any provider publishing OIDC discovery and JWKS:

LAW_MCP_AUTH_MODE=oauth \
LAW_MCP_AUTH_ISSUER=https://login.microsoftonline.com/<tenant>/v2.0 \
LAW_MCP_AUTH_AUDIENCE=api://law-scrapper \
LAW_MCP_AUTH_RESOURCE_SERVER_URL=https://mcp.example.com/mcp \
  law-scrapper

Reachable through a reverse proxy — LAW_MCP_ALLOWED_HOSTS and LAW_MCP_ALLOWED_ORIGINS default to loopback-only (F18): a request whose Host or Origin header doesn't match gets 421/403 from the SDK's DNS-rebinding protection, even with a valid token. This default is deliberate — widening the allowlist is only permitted once an auth mode is configured (enforced at startup, D6). A reverse proxy that preserves the original Host (nginx's proxy_set_header Host $host, Caddy's default) needs both variables set to the public name, for either auth mode above:

LAW_MCP_ALLOWED_HOSTS='mcp.example.com:*' \
LAW_MCP_ALLOWED_ORIGINS='https://mcp.example.com' \
  law-scrapper

Providers issuing opaque tokens (GitHub) are not supported — they would require RFC 7662 introspection. TLS termination stays with the reverse proxy. /health is intentionally unauthenticated so container healthchecks work; it exposes the server version and circuit-breaker state. It is exempt from the rate limiter only for loopback peers — a probe arriving from another host is metered like any other request.

Rate limiting is always on for HTTP: 60 requests per 60 s, burst 10. Behind a proxy, set LAW_MCP_TRUSTED_PROXIES (addresses or CIDRs) — otherwise X-Forwarded-For is ignored and every client shares one bucket.

Source modules

The 13 tools below are the core. The server can also load source modules: separately installed packages that add tools for further legal-information sources. Every installed module is active unless LAW_MCP_SOURCE_MODULES switches it off. The variable is a comma-separated list read from left to right, where the last token about a module wins: -name disables a module, name enables it, -* disables all. To run only woj, write LAW_MCP_SOURCE_MODULES=-*,woj; a bare woj leaves the other installed modules on. A disabled module is never imported, and a module that fails to load stops the server with a message naming it.

With no module installed nothing changes. Module tools are described by each module and are not part of the tool table below. A module can also serve further publisher codes through the tools below instead of adding tools; modules/voivodeship-journals does this for the voivodeship official journals (POL_WOJ_*, for example POL_WOJ_PM/2025/5051), see its README. Conventions, the full grammar and the gates: docs/source-modules.md.

Tools reference

Law Scrapper MCP provides 13 tools for legal research and analysis:

Every tool is read-only (readOnlyHint=true, destructiveHint=false, idempotentHint=true), and openWorldHint=true means the tool calls api.sejm.gov.pl, while false means it never does: it works on data already held in the server's memory or computes its answer locally.

1. get_system_metadata(category)

Retrieve system metadata for filtering and searching legal acts.

Parameters:

• category (string, default: "all") - Metadata category: "keywords", "publishers", "statuses", "types", "institutions", "coverage", or "all"

Returns: Keywords, publishers, document types, statuses, and institutions available in the system; with "coverage", what each active source covers (newest document, count, years)

Examples:

- Get all available search keywords
- Check what each source covers (newest document, years) - category "coverage"
- Retrieve all legal document types
- List all publishers (DU, MP)
- Get all document statuses
- Get complete system metadata

2. search_legal_acts(publisher, year, keywords, detail_level, status, type)

Search for legal acts with advanced filtering options.

Parameters:

• publisher (string) - Publisher code: "DU" (Dziennik Ustaw) or "MP" (Monitor Polski)

• year (integer) - Publication year (e.g., 2024)

• keywords (string) - Search keywords (AND logic - use multiple searches for OR)

• detail_level (string, default: "standard") - Response detail: "minimal", "standard", or "full"

• status (string, optional) - Document status filter

• type (string, optional) - Document type filter

Returns: List of matching legal acts with metadata

Search note: Multiple keywords use AND logic. Search one keyword at a time for OR behavior.

Examples:

- Search DU 2024 for "environment protection" acts
- Find all MP 2023 acts with status "active"
- Search for COVID-19 related legislation
- Find acts by specific type (e.g., "regulation")
- Get minimal detail results for quick scanning

3. browse_acts(publisher, year, detail_level)

Browse all legal acts published in a specific year by publisher.

Parameters:

• publisher (string) - Publisher code: "DU" or "MP"

• year (integer) - Publication year

• detail_level (string, default: "standard") - Response detail: "minimal", "standard", or "full"

Returns: Complete list of acts published in the specified year

Examples:

- Browse all DU acts from 2024
- Get minimal details of all MP acts from 2023
- Browse full details of DU 2022 legislation
- Get an overview of acts by publisher and year
- Track legislation published in a specific year

4. filter_results(result_set_id, pattern, field, type_equals, ...)

Filter and narrow down previously retrieved search/browse/changes results.

Parameters:

• result_set_id (string) - Result set ID from a previous search/browse/changes call (e.g., "rs_1")

• pattern (string, optional) - Regex pattern for text search (supports OR: "podatek|VAT|akcyza")

• field (string, default: "title") - Field to search: "title", "eli", "status", "type", "publisher"

• type_equals (string, optional) - Exact match on document type (e.g., "Ustawa", "Rozporządzenie")

• status_equals (string, optional) - Exact match on status (e.g., "akt obowiązujący", "akt uchylony")

• year_equals (integer, optional) - Exact match on publication year

• date_field (string, optional) - Date field for range filter: "promulgation_date" or "effective_date"

• date_from / date_to (string, optional) - Date range (YYYY-MM-DD)

• sort_by (string, optional) - Sort field: "title", "year", "pos", "promulgation_date", etc.

• sort_desc (boolean, default: false) - Sort descending

• limit (integer, optional) - Maximum results to return

Returns: Filtered results with a new result_set_id for chained filtering

Examples:

- Filter search results to only "Rozporządzenie" type
- Search titles with regex "zdrow|apteka|lekar"
- Filter by date range and sort by promulgation date
- Chain filters: first by type, then by regex pattern
- Get top 10 most recent results

Result-set scope. Every stored result set declares whether it is the whole answer (complete) or a window cut from a larger corpus (page), in the result_set_scope field of the response and in list_result_sets. The distinction matters because filter_results narrows the set, not the query: an empty filter over a page set means "no match among these twenty records", not "no such act exists". Responses say so explicitly — a filter that matches nothing over a window sets no_match_is_inconclusive. track_legal_changes is the one tool whose set is always complete, because it stores the whole fetched range.

5. get_act_details(eli, load_content, detail_level)

Retrieve detailed information about a specific legal act and optionally load its content.

Parameters:

• eli (string) - Act identifier in format "PUBLISHER/YEAR/NUMBER" (e.g., "DU/2024/1")

• load_content (boolean, default: false) - Load act content into Document Store for section reading

• detail_level (string, default: "standard") - Response detail: "minimal", "standard", or "full"

Returns: Act metadata (title, publication date, status, type, etc.), table of contents if load_content=true

content_status: Outcome of the content-loading half of this call, one of:

• not_requested - load_content was not set and the act was not already in the Document Store

• loaded - content is in the Document Store, ready for read_act_content / search_in_act

• unavailable - the act permanently has no readable text (no HTML/PDF source, a 404 on either text fetch, or an empty extraction from either format); retrying will not change that, and the server remembers the absence for the metadata cache TTL rather than asking the API again

A transient upstream failure during loading (an open circuit breaker, a timeout, an HTTP 5xx from api.sejm.gov.pl) is not a content_status value — it fails the call itself with isError=true, distinct from content_status="unavailable". Retrying a transient failure may succeed; retrying an unavailable act will not.

Examples:

- Get metadata for act DU/2024/1
- Load act content for section-level reading
- Get full details including table of contents
- Retrieve act status and publication information
- Load multiple acts for comparison

6. read_act_content(eli, section)

Read content from a specific section of a loaded legal act.

Parameters:

• eli (string) - Act identifier (must be loaded first via get_act_details with load_content=true)

• section (string) - Section to read (e.g., "Art. 1", "Chapter 2", "Preamble")

Returns: Content of the requested section

Workflow note: Must call get_act_details(eli="...", load_content=true) first, then use this tool.

Examples:

- Read Article 1 from loaded act
- Get Chapter 2 content
- Read the Preamble section
- Access specific numbered articles
- Navigate act by chapters

7. search_in_act(eli, query)

Search for specific terms within a loaded legal act.

Parameters:

• eli (string) - Act identifier (must be loaded first via get_act_details with load_content=true)

• query (string) - Search term or phrase

Returns: Matching sections with context and location

Examples:

- Find all mentions of "penalty" in loaded act
- Search for specific legal terms
- Locate articles containing "fine" or "punishment"
- Find definitional sections
- Search for specific references

8. analyze_act_relationships(eli, relationship_type)

Analyze legal relationships and references of an act (amendments, references, etc.).

Parameters:

• eli (string) - Act identifier

• relationship_type (string, default: "all") - Type: "amends", "amended_by", "references", "referenced_by", or "all"

Returns: List of related acts and their relationships

Examples:

- Find which acts amend this legislation
- See what acts this legislation amends
- Get all legal references in the act
- Find acts that reference this legislation
- Analyze complete act relationship network

9. track_legal_changes(date_from, date_to, publisher, keywords)

Track legal changes and new acts within a date range.

Parameters:

• date_from (string) - Start date (YYYY-MM-DD format)

• date_to (string) - End date (YYYY-MM-DD format)

• publisher (string, optional) - Filter by publisher: "DU" or "MP"

• keywords (string, optional) - Filter by keywords

Returns: Legal acts published in the date range

Examples:

- Track changes from 2024-01-01 to 2024-12-31
- Find new DU acts from last month
- Get changes published in past 7 days
- Track legislation on specific topics over time
- Monitor legal changes by publisher and date range

10. calculate_legal_date(days, months, years, base_date)

Calculate legal dates with intuitive sign convention.

Parameters:

• days (integer, default: 0) - Days offset (+future, -past)

• months (integer, default: 0) - Months offset (+future, -past)

• years (integer, default: 0) - Years offset (+future, -past)

• base_date (string, optional) - Base date (YYYY, YYYY-MM, or YYYY-MM-DD format, defaults to today)

Returns: Calculated date and relative description

Sign convention: Positive = future, Negative = past

Examples:

- Get current date (call with no parameters)
- Calculate date 30 days in the future (+30)
- Calculate date 6 months in the past (-6 months)
- Calculate date 1 year from a specific date
- Calculate legal deadlines and periods

11. compare_acts(eli_a, eli_b)

Compare metadata of two legal acts.

Parameters:

• eli_a (string) - ELI identifier of the first act (e.g., "DU/2024/1692")

• eli_b (string) - ELI identifier of the second act (e.g., "DU/2024/1716")

Returns: Comparison of titles, types, statuses, dates, keywords overlap and differences

Examples:

- Compare two acts from the same year
- Compare old and new versions of legislation
- Identify metadata differences between related acts

12. list_result_sets()

Display active result sets stored in memory.

Returns: List of result sets with IDs, query summaries, counts, and creation times

13. list_loaded_documents()

Display documents loaded into the Document Store.

Returns: List of loaded documents with ELIs, sizes, section counts, and timestamps

Document Store workflow

The Document Store pattern enables efficient content navigation and search within legal acts:

Workflow steps

1. Load an act - Call get_act_details(eli="DU/2024/1", load_content=true) to load the act into the Document Store

2. Read sections - Use read_act_content(eli="DU/2024/1", section="Art. 1") to read specific sections

3. Search within act - Use search_in_act(eli="DU/2024/1", query="penalty") to find terms

Benefits

• Efficient memory usage (configurable max documents and TTL)

• Fast section-level navigation without refetching

• Search within loaded acts without API calls

• Automatic content processing (PDF→text, HTML→Markdown)

Configuration

• LAW_MCP_DOC_STORE_MAX_DOCUMENTS - How many acts to keep in memory (default: 10)

• LAW_MCP_DOC_STORE_MAX_SIZE_BYTES - Maximum memory usage (default: 5 MB)

• LAW_MCP_DOC_STORE_TTL - How long to keep acts in memory (default: 2 hours)

Project structure

law-scrapper-mcp/
├── src/law_scrapper_mcp/
│   ├── __init__.py
│   ├── __main__.py              # Entry point for python -m
│   ├── server.py                # MCPServer, lifespan, transport config
│   ├── config.py                # Pydantic settings (env vars)
│   ├── logging_config.py        # Structured logging setup
│   ├── models/                  # Pydantic models
│   │   ├── enums.py            # Enumerations
│   │   ├── api_responses.py    # Sejm API response models
│   │   ├── tool_inputs.py      # Tool input models
│   │   └── tool_outputs.py     # Tool output models
│   ├── client/                  # HTTP client
│   │   ├── sejm_client.py      # AsyncClient with retry and circuit breaker
│   │   ├── cache.py            # Async TTL cache implementation
│   │   ├── circuit_breaker.py  # Circuit breaker for API protection
│   │   └── exceptions.py       # Custom exceptions (Polish messages)
│   ├── services/                # Business logic
│   │   ├── metadata_service.py    # Metadata retrieval
│   │   ├── search_service.py      # Search and browse
│   │   ├── act_service.py         # Act details and content
│   │   ├── changes_service.py     # Change tracking
│   │   ├── document_store.py      # In-memory act storage
│   │   ├── result_store.py        # Search result persistence and filtering
│   │   ├── content_processor.py   # PDF/HTML processing
│   │   └── response_enrichment.py # Response hints
│   └── tools/                   # MCP tool definitions
│       ├── annotations.py       # Shared ToolAnnotations presets (READ_ONLY_REMOTE / READ_ONLY_LOCAL)
│       ├── metadata.py          # get_system_metadata
│       ├── search.py            # search_legal_acts
│       ├── browse.py            # browse_acts
│       ├── act_details.py       # get_act_details
│       ├── act_content.py       # read_act_content
│       ├── act_search.py        # search_in_act
│       ├── relationships.py     # analyze_act_relationships
│       ├── filter_results.py    # filter_results, list_result_sets
│       ├── changes.py           # track_legal_changes
│       ├── compare.py           # compare_acts
│       ├── dates.py             # calculate_legal_date
│       └── error_handling.py    # Centralized @handle_tool_errors decorator
├── tests/
│   ├── unit/                    # Unit tests
│   └── integration/             # Integration tests with Sejm API
├── Dockerfile                   # Container image definition
├── docker-compose.yml           # Multi-service setup
├── pyproject.toml              # Project metadata and dependencies
├── uv.lock                      # Reproducible dependency lock
└── README.md                    # This file

Docker

Security and deployment

The server binds 127.0.0.1 by default; binding beyond loopback (as Docker port publishing requires) fails at startup unless an authentication mode is configured — see "Authenticated remote deployment" above. When exposing the HTTP transport (streamable-http) to a network, place the server behind a reverse proxy (nginx, Caddy, Traefik) with TLS termination — this project verifies bearer tokens and OAuth 2.1/OIDC access tokens, but does not terminate TLS itself. The /health endpoint is unauthenticated and intended for container healthchecks only — do not expose it publicly without access controls. Dependency versions are pinned in uv.lock with security overrides in pyproject.toml (cryptography, urllib3, idna, werkzeug, requests).

Host/Origin allowlist (DNS-rebinding protection): the official MCP SDK only auto-enables Host/Origin validation when the server binds to a literal loopback address (127.0.0.1, localhost, ::1). docker-compose.yml sets LAW_MCP_HOST=0.0.0.0 so Docker can publish the port, which would otherwise leave that validation disabled. server.py passes transport_security explicitly (build_transport_security()) so requests are still validated against the configured allowlist — Host outside it gets 421, Origin outside it gets 403 — independent of the bind address. The allowlist defaults to loopback only (127.0.0.1:* / localhost:* / [::1]:* and the matching http:// origins) and is widened through LAW_MCP_ALLOWED_HOSTS / LAW_MCP_ALLOWED_ORIGINS, which startup validation permits only once an authentication mode is configured. This restores the pre-3.0.0 FastMCP posture for /mcp. It is a defense-in-depth layer, not a substitute for the authentication mode required to bind beyond loopback in the first place.

One deliberate difference from the pre-3.0.0 server: the SDK applies this validation inside the Streamable HTTP app, not as whole-app middleware, so /health is not covered by the allowlist and answers any Host. That is what keeps container healthchecks working when they connect by container name or bridge IP, but it also means /health discloses the server name and version — and, since it now also reports the circuit breaker's circuit_state and failure_count, the health of the Sejm API integration — to anything that can reach the published port. FastMCP guarded /health too. Restrict the published port, or front it with a proxy, if that disclosure matters to you.

Health contract: the response body carries an upstream object. upstream.circuit_state is closed, open, half_open, or unknown before the lifespan has started (or after it has been torn down) — in that unknown case upstream.failure_count is omitted entirely; otherwise it sits alongside circuit_state. /health still answers 200 while the breaker is open: restarting the container cannot repair an outage of api.sejm.gov.pl, and under restart: unless-stopped a 503 would turn someone else's outage into a restart loop.

Dockerfile

The included Dockerfile builds a containerized Law Scrapper MCP server:

FROM python:3.13-slim
WORKDIR /app
COPY . .
RUN pip install -e .
EXPOSE 7683
CMD ["law-scrapper"]

Build and run:

# Build the image
docker build -t law-scrapper-mcp .

# Run with STDIO transport
docker run -it law-scrapper-mcp

# Run with HTTP transport, published on the host — the image no longer
# defaults LAW_MCP_HOST to 0.0.0.0, so publishing the port beyond loopback
# requires setting the host explicitly together with an authentication mode
docker run -it -p 7683:7683 \
  -e LAW_MCP_TRANSPORT=streamable-http \
  -e LAW_MCP_HOST=0.0.0.0 \
  -e LAW_MCP_AUTH_MODE=bearer \
  -e LAW_MCP_AUTH_TOKEN="$(openssl rand -base64 32)" \
  -e LAW_MCP_LOG_LEVEL=DEBUG \
  law-scrapper-mcp

docker-compose.yml

Deployment with docker-compose:

# LAW_MCP_AUTH_TOKEN is required — docker-compose.yml fails fast without it
export LAW_MCP_AUTH_TOKEN=$(openssl rand -base64 32)

# Start service
docker compose up -d

# View logs
docker compose logs -f

# Stop service
docker compose down

Migration guide (v1 to v2)

If upgrading from v1.0.2, note these breaking changes:

What's new in v3.0.0

• Official Python MCP SDK — Replaced the prior third-party framework with mcp[cli]==2.0.0 and MCPServer[AppContext]

• Native structured responses — Tools return EnrichedResponse with outputSchema and object structuredContent (no JSON strings)

• Protocol-visible errors — Tool failures surface as isError=true instead of in-body error fields

• Stateless Streamable HTTP — /mcp endpoint only; legacy event-stream transport removed

• Pagination — PageInfo model exposed as the page_info field with limit/offset on search, browse, metadata, changes, TOC, sections, and in-act search

• Typed AppContext — Lifespan resources accessed via ctx.request_context.lifespan_context

• Domain services — ComparisonService, RelationshipService, and DateService extracted from tool adapters

• Transport tests — Real STDIO subprocess, loopback HTTP, and MCP conformance in CI

• Explicit Host/Origin allowlist on HTTP — transport_security is passed explicitly so DNS-rebinding protection applies regardless of the configured bind host; see Security and deployment

Migration guide (v2 to v3)

What's new in v2.4.0

• Security hardening — Framework upgrade and dependency overrides close 51 Dependabot alerts (cryptography, urllib3, pillow, starlette, and others)

• Framework 3.x — Lifespan context API, app.run() for HTTP transport, @custom_route for /health

• Integration tests — In-memory client tests for core tools (metadata, search, dates, act details)

• CI and Dependabot — Automated quality gates and weekly dependency updates

What's new in v2.3.1

• uvx / tool registration fix — Fixed NameError: name 'Annotated' is not defined when running via uvx --from "git+https://github.com/numikel/law-scrapper-mcp" law-scrapper. Removed from __future__ import annotations from compare.py so parameter type hints resolve correctly during tool registration.

What's new in v2.3.0

• 3 new tools — compare_acts, list_result_sets, list_loaded_documents (total: 13 tools)

• Circuit breaker — Protects against cascading failures when Sejm API is unavailable

• Centralized error handling — @handle_tool_errors decorator with error classification and full tracebacks

• asyncio.Lock migration — All stores use asyncio.Lock for proper async compatibility

• Default search limit — Search/browse return max 20 results by default to limit token usage

• Health endpoint — /health for Docker deployments with streamable-http transport

• Polish error messages — All exception messages in Polish for consistent user experience

• Decision tree docstrings — "When to use" / "When NOT to use" for all tools

Development

Setup

# Install dependencies
uv sync

# Install with dev dependencies
uv sync --extra dev

Running tests

# Run unit tests
uv run pytest tests/unit/ -v

# Run integration tests (mocked Sejm API via respx; no live network required)
uv run pytest tests/integration/ -v -m integration

# Run all tests with coverage
uv run pytest --cov=law_scrapper_mcp --cov-report=term-missing

# Run with timeout for slow tests
uv run pytest --timeout=10 -v

# Audit the locked dependencies for known vulnerabilities (same check as the CI audit job)
uv run pip-audit

MCP integration testing

Law Scrapper MCP uses three automated test layers to verify all 13 tools and their MCP protocol interaction:

Layer 1 — unit tests (tests/unit/):

uv run pytest tests/unit/ -v

Layer 2 — in-memory integration (official mcp.Client, mocked Sejm API via respx):

uv run pytest tests/integration/test_tools_e2e.py \
  tests/integration/test_content_pagination.py \
  tests/integration/test_result_pagination.py -v -m integration

Layer 3 — transport integration (real STDIO subprocess and loopback Streamable HTTP):

uv run pytest tests/integration/test_stdio_transport.py \
  tests/integration/test_http_transport.py -v -m integration

Full integration suite:

uv run pytest tests/integration/ -v -m integration

What the tests verify:

• All 13 tools registered with preserved public arguments

• Native structuredContent payloads with outputSchema (not JSON strings)

• isError=true on tool execution failures

• page_info (PageInfo) pagination metadata on paginated outputs

• Stateful workflows: search → filter → load act → read sections → search in content

• Real STDIO and HTTP transports against a running server process

Architecture:

• Test framework: pytest with asyncio (unit) and anyio (MCP client tests)

• HTTP mocking: respx intercepts httpx calls in in-memory integration tests

• MCP protocol: Official mcp.Client from the Python MCP SDK

• Helpers: tests/mcp_helpers.py (parse_tool_result for structuredContent assertions)

• CI: Legacy dependency gate, protocol transport tests, and MCP conformance against /mcp

See tests/TEST_SUITE_SUMMARY.md for the full file layout and patterns.

Code quality

The project follows MCP SDK best practices:

• Modular architecture - Separated concerns (models, client, services, tools)

• Type hints - Full type annotation with Pydantic models

• Async throughout - Async/await for all I/O operations

• Comprehensive examples - Minimum 5 examples per tool

• Tagged tools - Organized by category for easy discovery

• Annotated parameters - Clear descriptions for all inputs

• Structured logging - Configurable JSON/text formats

Running the server

# STDIO transport (default)
uv run python -m law_scrapper_mcp

# HTTP transport
LAW_MCP_TRANSPORT=streamable-http uv run python -m law_scrapper_mcp

# With debug logging
LAW_MCP_LOG_LEVEL=DEBUG uv run python -m law_scrapper_mcp

Contributing

1. Fork the repository

2. Create your feature branch (git checkout -b feature/amazing-feature)

3. Commit your changes using Conventional Commits format

4. Add tests for new functionality

5. Ensure all tests pass and coverage is maintained

6. Push to the branch (git push origin feature/amazing-feature)

7. Open a Pull Request

Development guidelines

• Follow MCP SDK patterns for tool definitions and structured output

• Include comprehensive examples and parameter descriptions

• Add appropriate tags for tool categorization

• Write async code throughout

• Add tests for all new functionality

• Update CHANGELOG.md with your changes

• Use English for all code comments and documentation

License

This project is licensed under the MIT License. See the LICENSE file for details.

Author

@numikel

Developed with help from:

And with models:

Legal disclaimer: This tool provides access to Polish legal documents for research purposes. Always consult with qualified legal professionals for legal advice and interpretation of laws.