rejstrik-mcp 0.8.0

Czech registry MCP that reads the documents

rejstrik-mcp

Add the Czech business registry to your Claude in 30 seconds — no API key. It reads the actual filed PDFs with your own subscription.

claude mcp add rejstrik -- uvx rejstrik-mcp

See it work

docs/media/cli-demo.txt is a real, keyless transcript — no API key set — resolving a company via ARES and listing its filed financial statements straight from Sbírka listin. A Claude Desktop screenshot of the rendered report card is still pending a manual capture (see docs/media/README.md).

Then ask: "What happened to Budějovický Budvar's finances last year?" — your agent resolves the company (ARES), pulls the filed statement PDF from the Sbírka listin, reads it itself, and gets deterministic ratios, red flags, and trends back from the server. No OCR pipeline, no server-side AI key, no scraping middleman.

Why this one

	agent-native (MCP)	reads filed PDFs	free & open source	works without any API key
cz-agents-mcp and similar	✅	❌	✅	✅
chytryrejstrik.cz	❌	partly (paid)	❌	—
rejstrik-mcp	✅	✅	✅	✅

Install

Claude Code: claude mcp add rejstrik -- uvx rejstrik-mcp

Claude Desktop: download rejstrik-mcp.mcpb from the latest GitHub release and double-click it (requires uv) — or add to claude_desktop_config.json:

{ "mcpServers": { "rejstrik": { "command": "uvx", "args": ["rejstrik-mcp"] } } }

Codex (~/.codex/config.toml):

[mcp_servers.rejstrik]
command = "uvx"
args = ["rejstrik-mcp"]

Any HTTP host: uvx rejstrik-mcp --http serves streamable HTTP on http://127.0.0.1:8000/mcp.

The tools

Your agent does the reading with your existing subscription; the server does everything deterministic:

Tool	What it does
find_company	Resolve a company by name or IČO (ARES)
list_filings	List Sbírka listin documents, financial statements first
get_filing	Download a statement PDF (latest, by year, or by id) — returns local path + page_count; embed controls whether the PDF bytes are also returned ("auto" default, "always", or "never")
read_filing_text	Keyless extraction of the PDF text layer for a page range (no LLM/OCR); pages without a text layer are reported honestly
analyze_financials	Your extracted figures in → ratios, red flags, IN05 distress index, year-over-year trends out (no LLM)
estimate_valuation	Your extracted figures in → indicative valuation range (book value, capitalized earnings, multiples), no LLM. Not investment advice
render_card	The report as a card — an interactive HTML card for MCP Apps hosts, a markdown summary for text-only hosts like Claude Code
check_insolvency	Insolvency register (ISIR)
get_statutory_bodies	Directors / statutory bodies (ARES)
check_vat	VAT registration + unreliable-payer flag (ARES + ADIS)
get_subsidies	State subsidies received (IS ReD / former CEDR)
get_contracts	Public contracts involving the company (Registr smluv)

Beneficial owners. The public part of ESM (Evidence skutečných majitelů) closed on 2025-12-17 following an EU Court of Justice ruling, so beneficial-owner lookups are intentionally not offered here — a documented scope decision, not a gap.

Use the built-in analyze-company prompt (shows up as a slash command in Claude) to run the whole loop — find → fetch PDFs → extract → analyze → card — including multi-year trends.

How it works

core/      shared HTTP + text utilities
registry/  ARES, ISIR (insolvency), ADIS (VAT), statutory bodies
filings/   verejnerejstriky.msp.gov.cz Sbirka listin client
           (falls back to legacy or.justice.cz when the new portal is blocked)
documents/ native-PDF extraction + document Q&A
analysis/  normalize -> ratios -> red flags -> trends (pure, no I/O)
service/   orchestration (registry + filings + documents + analysis)
cli/ mcp/  two faces over one core

The document engine sends PDFs directly to the configured model, so there is no OCR pipeline or vector database. Structured extraction uses Pydantic schemas; open-ended Q&A uses the same document abstraction and returns citations when the provider supports them.

A Note On Real-World Drift

Midway through the build, the Czech Ministry of Justice migrated the Sbirka listin from or.justice.cz to a new Nuxt portal (verejnerejstriky.msp.gov.cz). The filings client was re-pointed at the new portal's API. Registry, filings, insolvency, statutory-body, VAT, and ADIS lookups are covered by fixtures/unit tests; live smoke testing verified the registry/document analysis path against Budejovicky Budvar.

In July 2026 the new portal began returning Azure Front Door block responses — 403/429/5xx and 200-with-challenge-HTML interstitials — to automated clients. The filings client now treats all of these as block-shaped and falls back to the legacy or.justice.cz portal, so a single blocked edge does not break lookups. A scripts/smoke.py canary hits both portals directly and prints PASS/BLOCKED per endpoint, so this drift is caught before a release tag rather than in the field.

Industry valuation multiples data

src/rejstrik/analysis/data/industry_multiples.json vendors Damodaran Europe industry EV/EBITDA multiples (source, source_url, as_of and region are recorded in the file). NACE is used only as a mapping key into Damodaran's industry taxonomy — no hand-tuned multiples. Regenerate (network, manual, never in CI) with:

pip install xlrd
python scripts/import_damodaran_multiples.py --as-of YYYY-MM-DD

CI

The GitHub Actions workflow in .github/workflows/ci.yml runs ruff and pytest on Python 3.11 and 3.12. The test suite is deliberately offline and key-free, so a green CI run means the fixtures, parsers, service layer, CLI, and MCP registrations are internally consistent without live endpoint luck.

Development

pip install -e ".[dev]"
ruff check src/ tests/
ruff format --check src/ tests/
python -m pytest -q

Autopilot (unattended roadmap execution)

The July 2026 roadmap stages (docs/superpowers/specs/2026-07-13-*.md) can be run unattended: Opus writes each stage's implementation plan, Sonnet executes it in an isolated git worktree, and the script verifies (ruff + pytest) and merges only on green. State is derived from git, so the script can be stopped and rerun anywhere after a clone. Requires a logged-in Claude Code CLI; see docs/superpowers/specs/2026-07-13-autopilot-design.md.

python scripts/autopilot.py --dry-run   # show what would happen
python scripts/autopilot.py             # run remaining stages

Useful manual smoke tests:

rejstrik analyze "Budejovicky Budvar"
rejstrik-mcp
python scripts/smoke.py            # live network required — run before releases, not in CI

Attribution

The insolvency (ISIR), VAT/unreliable-payer (ADIS), and statutory-body registry clients are adapted from cz-agents-mcp (MIT, Martin Havel). See LICENSES/cz-agents-mcp-LICENSE.

Releasing

1. One-time: on pypi.org, add a Trusted Publisher for this GitHub repo (workflow release.yml, environment pypi).
2. Bump version in all four places so they agree: pyproject.toml, server.json (top-level and packages[0].version), mcpb/manifest.json, and src/rejstrik/__init__.py (__version__). tests/test_version_sync.py fails if any of them drift.
3. If the release changes the published server, re-run the MCP registry publisher flow with the updated server.json.
4. Commit, tag vX.Y.Z, push the tag. CI builds, publishes to PyPI, and attaches artifacts to the GitHub release.

License

MIT.