prawo-pl-mcp

One MCP server for Polish legal data.

Ten independent connectors - case law, legislation, the company register, tax rulings, public procurement, data-protection decisions, plus EU extras - behind five tools. Each connector stays a separate package with its own repository; this server spawns them on demand and translates between their conventions.

"Prawo" is Polish for "law"; the name follows yargi-mcp's move of naming a country server in its own language.

Why an aggregator

The connectors cover Polish legal data well, but as separate MCP servers, one per database - ten installs and ten config entries before the first question. yargi-mcp showed the fix for Turkish law: 16 institutions, one server. prawo-pl-mcp does the same for Poland:

• Five tools instead of 42. pl_list_sources, pl_search, pl_get_document, pl_call and pl_coverage - fewer tool schemas means less context spent before the LLM starts working.

• One convention. The connectors grew organically: dateFrom here, date_from there, pages counted from 0 or from 1 depending on the repo. Here page is always 1-based and dates are always date_from/date_to (YYYY-MM-DD); the translation happens inside. One name per concept applies to the aggregator's own tools too: page is the page parameter of both pl_search and pl_get_document (page_number still works as a deprecated alias), and pl_call takes arguments or its alias args.

• Native names on the table. Every catalog entry carries native_params - the native parameter name each unified one maps to for that source, so article_number (eu-compliance) is visible without spawning the connector or reading an upstream error. pl_call checks the names it was given against the connector's own schema and says which one it meant.

• Nothing preinstalled. Connectors run as subprocesses via npx -y / uvx, downloaded on the first call to a given source and kept alive afterwards. A source that cannot start reports a readable error while the rest keep working.

• Paginated documents. Full judgment and act texts are chunked at ~5000 characters per page (yargi-mcp's pattern), so a 200-page ruling does not flood the context window. Sources that paginate server-side (isap) get the page forwarded natively and are reported as pagination: "native" - the aggregator does not re-cut an already-cut page and does not invent a total_pages it cannot know.

Related MCP server: mcp-nsa

Coverage

Polish sources are tagged group: pl, EU extras group: eu - filter with pl_list_sources(group="pl").

Tools

The typical flow an LLM follows (spelled out in the server's instructions): pl_list_sources → pl_search → pl_get_document, with pl_call for the specialized tools each connector brings.

Install

Requires Python ≥ 3.11 plus the runtimes of the sources you use: Node.js ≥ 18 for npx sources, uv for uvx sources. If a runtime is missing, its sources report source_unavailable and the rest work.

Claude Code

claude mcp add pl-legal -- uvx prawo-pl-mcp

Claude Desktop / any MCP client (stdio)

{
  "mcpServers": {
    "pl-legal": {
      "command": "uvx",
      "args": ["prawo-pl-mcp"]
    }
  }
}

Remote (Streamable HTTP)

uvicorn prawo_pl_mcp.asgi:app --host 0.0.0.0 --port 8000

Open by default (public, read-only data). Set PRAWO_PL_MCP_API_KEY to require X-API-Key: <key> or Authorization: Bearer <key> on every request.

Configuration

Architecture

The aggregator is a thin proxy (ADR 0001). Connectors are not imported, vendored or forked; the aggregator speaks MCP to them over stdio the same way any client would. The whole layer is a source registry (one dataclass entry per connector), a lazy subprocess pool, a parameter translator and a paginator. Adding a source means adding a registry entry.

LLM client ──MCP──▶ prawo-pl-mcp ──MCP/stdio──▶ npx @matematicsolutions/mcp-saos
                        │         ──MCP/stdio──▶ uvx kio-orzeczenia-mcp
                        │         ──MCP/stdio──▶ ... (spawned on first use)
                        └─ registry + param mapping + 5000-char pagination

Every call lands in a JSONL audit log (timestamp, tool, source, parameter hash, latency - never document content).

Development

git clone https://github.com/matematicsolutions/prawo-pl-mcp && cd prawo-pl-mcp
uv sync --extra dev
uv run pytest          # offline tests: registry, dispatch mapping, pagination, drift
uv run prawo-pl-mcp    # stdio server

License

Apache-2.0. Individual connectors carry their own licenses (MIT or Apache-2.0), their own rate limits and their own terms toward upstream databases - the aggregator adds no caching and no transformation beyond pagination, so each connector's constraints apply unchanged.