Czech Law MCP

Czech Law MCP

▶ Try this MCP instantly via Ansvar Gateway

50 free queries/day · no card required · OAuth signup at ansvar.eu/gateway

One endpoint, one OAuth signup, access from any MCP-compatible client.

Connect

Claude Code (one line):

claude mcp add ansvar --transport http https://gateway.ansvar.eu/mcp

Claude Desktop / Cursor — add to claude_desktop_config.json (or mcp.json):

{
  "mcpServers": {
    "ansvar": {
      "type": "url",
      "url": "https://gateway.ansvar.eu/mcp"
    }
  }
}

Claude.ai — Settings → Connectors → Add custom connector → paste https://gateway.ansvar.eu/mcp

First request opens an OAuth flow at ansvar.eu/gateway. After signup, your client is bound to your account; tier (free / premium / team / company) determines fan-out, quota, and which downstream MCPs are reachable.

Self-host this MCP

You can also clone this repo and build the corpus yourself. The schema,fetcher, and tool implementations all live here. What is not in the repo isthe pre-built database — TDM and standards-licensing constraints on theupstream sources mean we host the corpus on Ansvar infrastructure ratherthan redistribute it as a public artifact.

Build your own: run this repo's ingestion script (entry-point varies perrepo — typically scripts/ingest.sh, npm run ingest, or make ingest;check the repo root).

MCP server for Czech Law — 45,899 statutes from www.e-sbirka.cz.

What this is

This server indexes the legal materials listed under Sources below andexposes them via the Model Context Protocol. Part of the Ansvar MCP fleet —source-available servers published for self-hosting.

It makes no outbound network calls except to the upstream sources duringingestion — no analytics, no phone-home.

Coverage

• Corpus: Czech Law — 45,899 statutes, 461,231 provisions
• Jurisdiction code: CZ
• Corpus snapshot: 2026-02-22

The corpus is rebuilt from the upstream sources by the included ingestion script; re-run periodically to refresh.

See Sources below for source URLs, terms, and reuse conditions.

Why this exists

LLMs answering compliance, security, or legal questions from training dataalone will fabricate citations — confidently producing article numbers,statute names, and source URLs that do not exist, or that do not say whatthe model claims. This MCP exists so an agent can call a tool that returnsthe real text, the real identifier, and the real source URL straight fromthe indexed materials — and ground an answer rather than recall it.

One MCP, one corpus. The point is composition.

The Ansvar Gateway (ansvar.eu) joins this MCPwith the rest of the Ansvar fleet behind a single authenticatedendpoint — 300+ servers covering legal jurisdictions, EUregulations, security frameworks, sector regulators, privacy-patterncatalogues, and risk-scoring tools. That lets an agent run cross-domainworkflows that no single MCP can serve alone:

• Threat model and TARA. Threat enumeration → known componentvulnerabilities → severity scoring → applicable AI, cybersecurity, andautomotive obligations → privacy threats. Every finding traceable toits source.
• Gap analysis. Target framework requirements → current-stateevidence → unmet obligations → remediation guidance and authorityopinions. Every gap traceable to the specific requirement that flaggedit.
• Data Protection Impact Assessment. Privacy regulation articles →national DPA guidance → privacy-pattern catalogue → applicable caselaw.

Getting high-quality citations

Citation accuracy degrades when an agent's context fills up. Long inputscause retrieval-stage drift — the model recalls claim text correctly butmisattributes the source. Two practices keep accuracy high:

1. Focused first pass, checking-agent second pass. Query a small,relevant set of MCPs first, then run a separate agent whose only jobis to re-resolve each citation against the source MCP and flag anythat no longer match. The checking agent uses the same MCP tools asthe synthesis agent.
2. Pull the source text verbatim when in doubt. Every citation anagent emits points back to a tool call against this server. You — oranother agent — can call the same tool with the same identifier andread the raw statute, article, or standard text directly. If theverbatim text doesn't support what the agent claimed, the citationwas misused, regardless of whether the identifier was real.

Both patterns work the same way self-hosted or through the gateway.

Two ways to use it

Self-host (free, Apache 2.0) — clone this repo, run the ingestion script to build your local database from the listed upstream sources, point your MCP client at the local server. Instructions below.

Use the hosted gateway — for production use against the curated,kept-fresh corpus across the full Ansvar MCP fleet, with citation enrichmentand multi-jurisdiction fan-out — see ansvar.eu.

Self-hosting

Install

git clone https://github.com/Ansvar-Systems/Czech-law-mcp.git
cd Czech-law-mcp
npm install

Build the database

npm run build:db

Ingestion is a snapshot — your local copy goes stale until you re-run it. The hosted gateway corpus is refreshed continuously.

Premium tier (case-law)

The premium tier extends the free corpus with curated Supreme Court decisions from the Sbírka soudních rozhodnutí a stanovisek published at sbirka.nsoud.cz. License basis is the same Czech-Statutory-PD §3 carve-out that covers statutes — court decisions are excluded from copyright protection under Czech Copyright Act 121/2000 Coll.

Build the premium database in two steps. First, ingest the case-law corpus (single-threaded, ≥ 2 seconds per request, honouring robots.txt):

npm run ingest:nsoud-sbirka                  # full corpus (~9,790 decisions, ~6 hours)
npm run ingest:nsoud-sbirka -- --limit 100   # smoke test (200 seconds)
npm run ingest:nsoud-sbirka -- --resume      # skip already-ingested URLs

Seed files land under data/case-law-seed/{ns-sbirka-<id>}.json. Then build the premium database, which combines the e-Sbírka statute corpus with the case-law seeds:

npm run build:db:premium

The premium build writes to data/database-premium.db (~1.3 GB statutes + case-law growth). Case-law rows are stored in the same legal_documents table with type='case_law', publisher='Nejvyšší soud', license='Czech-Statutory-PD'. Cited statutes parsed from each decision's Předpisy field land in cross_references.

The full Supreme Court decision search at rozhodnuti.nsoud.cz is out of scope — that subdomain blocks all non-DG_JUSTICE_CRAWLER user-agents via robots.txt and requires a separate written exception from NS IT. The Sbírka collection is the curated, publication-grade subset.

Configure your MCP client

{
  "mcpServers": {
    "czech-law-mcp": {
      "command": "node",
      "args": ["dist/index.js"]
    }
  }
}

Sources

What this repository does not provide

This repository's source — the MCP server code, schema, and ingestion script — is licensed under Apache2.0. The license below covers the code in this repository only; it does notextend to the upstream legal materials.

Running ingestion may download, cache, transform, and index materials from the listed upstream sources. You are responsible for confirming that your use of those materials complies with the source terms, attribution requirements, robots/rate limits, database rights, copyright rules, and any commercial-use or redistribution limits that apply in your jurisdiction.

License

Apache 2.0 — see LICENSE. Commercial use, modification, andredistribution of the source code in this repository are permitted underthat license. The license does not extend to upstream legal materials downloaded by the ingestion script; those remain governed by the source jurisdictions' own publishing terms (see Sources above).

The Ansvar gateway

If you'd rather not self-host, ansvar.eu provides thisMCP plus the full Ansvar fleet through a single authenticated endpoint, withthe curated production corpus, multi-MCP query orchestration, and citationenrichment.

Issues: github.com/Ansvar-Systems/Czech-law-mcp/issues · Security: [email protected]