PolishOpenData

.NET libraries and an MCP server for Polish public registers — company lookup in KRS, the VAT whitelist (Biała Lista) with bank-account checks (online, or offline against the Ministry of Finance flat file), a client for the GUS REGON register (library only), cadastral parcel lookup (ULDK) and address geocoding (UUG) from GUGiK, and offline validation of NIP, REGON, KRS, NRB and land-registry (KW) numbers.

🇵🇱 Polska wersja

Unofficial project. Not affiliated with the Ministry of Justice, the Ministry of Finance or any other public body. Results are informational, not legal or tax advice; the registers are the authoritative source.

Packages

Offline validation

No network, no DI:

dotnet add package PolishOpenData.Core

using PolishOpenData;

Console.WriteLine(Nip.TryParse("774-000-14-54", out var nip) ? $"Valid NIP {nip}" : "Invalid NIP"); // Valid NIP 7740001454
Console.WriteLine(KwNumber.TryParse("WA1M/00012345/1", out var kw) ? kw.Court?.Name : "Invalid KW number"); // Sąd Rejonowy dla Warszawy-Mokotowa …
Console.WriteLine(string.Join(", ", KwNumber.SuggestCorrections("WAIM/00012345/1"))); // WA1M/00012345/1

Quick start: check a counterparty

dotnet add package PolishOpenData.BialaLista
dotnet add package PolishOpenData.Krs
dotnet add package Microsoft.Extensions.Http.Resilience

using Microsoft.Extensions.DependencyInjection;
using PolishOpenData;
using PolishOpenData.BialaLista;
using PolishOpenData.Krs;
using System.Threading.Tasks;

var services = new ServiceCollection();
services.AddBialaListaClient(o => o.TrackQuota = true)
    .AddStandardResilienceHandler(o => o.Retry.ShouldHandle = _ => new ValueTask<bool>(false));
services.AddKrsClient().AddStandardResilienceHandler();
using var provider = services.BuildServiceProvider();

var vat = provider.GetRequiredService<IBialaListaClient>();
var nip = Nip.Parse("774-000-14-54");

var result = await vat.FindByNipAsync(nip);
Console.WriteLine($"{result.Value?.Name}: {result.Value?.VatStatus} (request {result.RequestId})");

var check = await vat.CheckBankAccountAsync(nip, Nrb.Parse("06 1600 1127 1843 9838 2000 0034"));
Console.WriteLine(check.Value ? "Account is on the whitelist" : "Account is NOT on the whitelist");

var krs = provider.GetRequiredService<IKrsClient>();
var extract = await krs.GetCurrentExtractAsync(result.Value!.Krs!.Value);
Console.WriteLine(extract.Extract?.ToSummary().ShareCapital);

Retries are off for Biała Lista: every retry the resilience handler would make is another upstream request that the quota guard above cannot count, and once the daily limit is reached Biała Lista blocks the whole IP address — not just this process — until midnight. Register AddBialaListaClient once per process and reuse the resulting client (or pass one shared BialaListaQuotaTracker explicitly); a client created without DI and without a shared tracker gets its own private tracker that does not see requests made by other instances.

A runnable version is in samples/CheckCounterparty.

Keep RequestId. Every BialaListaResult carries the request identifier and time returned by Biała Lista (RequestId, RequestDateTime); keep them with your payment records to document when you checked the whitelist and what it answered.

Source-generated JSON. The KRS models use public converters from PolishOpenData.Krs.Serialization (KrsDateJsonConverter, KrsTimestampJsonConverter), so on .NET 8 or later you can add KrsCurrentExtract or KrsFullExtract to your own JsonSerializerContext (for example to cache extracts) without registering any converter. If you parse raw KRS API responses yourself instead of calling IKrsClient, note two things: the extract is wrapped in an odpis object, so deserialize that inner object; and the registry's field names (such as dataRejestracjiWKRS) match the properties only case-insensitively, so use JsonSerializerDefaults.Web — [JsonSourceGenerationOptions(JsonSerializerDefaults.Web)] on your context, or new JsonSerializerOptions(JsonSerializerDefaults.Web) with the reflection-based serializer on .NET Framework. Otherwise the properties stay null without an error.

Quick start: look up an entity in the GUS REGON register

PolishOpenData.Regon is a client for GUS's BIR1.2 web service: search by NIP, REGON or KRS number, batches of up to 20 identifiers, and the full reports of legal persons, sole traders and local units. It needs a key of your own (free: write to regon_bir@stat.gov.pl); GUS's public test environment works without a key but serves an old, anonymised copy of the register.

dotnet add package PolishOpenData.Regon

using Microsoft.Extensions.DependencyInjection;
using PolishOpenData;
using PolishOpenData.Bir;
using System.Linq;

var services = new ServiceCollection();
services.AddRegonClient(o => o.UserKey = Environment.GetEnvironmentVariable("MY_APP_REGON_KEY")); // your key, from your own configuration
// GUS's public test environment needs no key: services.AddRegonClient(o => o.Environment = RegonEnvironment.Test);
using var provider = services.BuildServiceProvider();
var regon = provider.GetRequiredService<IRegonClient>();

var found = await regon.FindByNipAsync(Nip.Parse("526-104-08-28"));
var entity = found.Value[0];
Console.WriteLine($"{entity.Name}, {entity.Address.City} (REGON {entity.Regon})");

var details = (await regon.GetDetailsAsync(entity, RegonInclude.Pkd)).Value[0];
Console.WriteLine($"{details.LegalFormName}; main PKD {details.Pkd!.First(p => p.IsPrimary).Code}");
Console.WriteLine($"Source: GUS REGON ({found.Environment}), retrieved {found.RetrievedAt:u}");

await regon.LogoutAsync();

The client limits itself to GUS's strictest window by default (3 calls per second, 120 per minute, 6,000 per hour; a batch of up to 20 identifiers is one call for the per-second limit, while the per-minute and per-hour limits count requests and identifiers: each identifier of a search counts one, and so does every other request — login, report, GetDataDateAsync, logout) and keeps one login session. A batch method checks its whole batch before the first request and throws QuotaExceededException without sending anything when it does not fit into what is left of the minute or hour window: under the default window one call carries at most 119 distinct identifiers (120 when the session is already logged in, fewer when other calls used the last minute). Whether GUS also counts every identifier of a batch toward its own per-second limit is not confirmed; if you send many batches, use smaller ones or space them out. Register AddRegonClient once per process and reuse the client: GUS counts requests per application instance, and a session works only from the IP address that created it. Call LogoutAsync when your process stops; the client never does it by itself. A client created without DI gets a private session and limiter.

Retries stay off for REGON, as for Biała Lista: a retry made by a resilience handler is another GUS request that the request limiter cannot count, and a SOAP fault, which means your request was wrong, arrives as HTTP 500 and would be retried. If you add AddStandardResilienceHandler, switch its retries off as in the Biała Lista example above.

Keys and the test environment. The library is bring-your-own-key: it ships no production key, never reads environment variables and never logs or prints your key. GUS's technical manual (section 3.6.3) calls it unacceptable for software to make every end user apply for their own key, but GUS told the maintainer on 2026-10-15 that for this project each end user may apply for and use their own key; that answer covers this project, so check GUS's rules before you distribute your own application built on the library. RegonEnvironment.Test selects GUS's public test service and its published test key (a UserKey that you configure is ignored there and never sent to the test service, so you can keep your own key configured and switch the environment); every result says which environment it came from (RegonResult.Environment), so anonymised test data (mostly from 2014) cannot pass for current data.

Cost. Details cost requests: a legal person takes a login, a search and one report (3 requests), each optional list (RegonInclude.Pkd, LocalUnits, Partners) one more, and a sole trader the general report plus one report per activity type. RegonResult.DataAsOf is filled once GetDataDateAsync has been called (one request).

Offline VAT whitelist check

The Ministry of Finance publishes the VAT whitelist every day as a flat file (plik płaski) of hashed NIP and account pairs. PolishOpenData.BialaLista.Offline downloads the file for a date, builds a local index once, and then answers any number of checks from disk: no request limit, and nothing about the checked NIP/account pairs leaves your machine. Use it for large batches (a payment run, an ERP import); use PolishOpenData.BialaLista when you need a request ID as proof.

dotnet add package PolishOpenData.BialaLista.Offline

using Microsoft.Extensions.DependencyInjection;
using PolishOpenData;
using PolishOpenData.BialaLista.Offline;

var services = new ServiceCollection();
services.AddVatWhitelistOffline(o => o.DatesToKeep = 3);
using var provider = services.BuildServiceProvider();
var offline = provider.GetRequiredService<IVatWhitelistOffline>();

await offline.PrepareAsync(); // optional: downloads and indexes today's file now instead of in the first check
var checks = await offline.CheckManyAsync(new[]
{
    (Nip.Parse("774-000-14-54"), Nrb.Parse("06 1600 1127 1843 9838 2000 0034")),
});
foreach (var c in checks)
    Console.WriteLine($"{c.Nip}: {c.Status} (file of {c.FileDate:yyyy-MM-dd}, generated {c.GeneratedOn:yyyy-MM-dd})");

• Three answers. Active, Exempt or NotFound. NotFound means not found in the file (the plain hash and every virtual-account mask of the account's bank were tried), not proof that the account is absent from the whitelist; MatchedVirtualAccountMask tells you when a pair was found through a mask.
• First use of a date is slow. It downloads about 250 MB and builds an index. That took about 4 minutes on a fast connection (October 2026), but much longer when the Ministry's server is slow, which happens during the working day. The download is deleted once the index is built. Call PrepareAsync ahead of time to avoid the wait: the Ministry publishes the file for day D at about 23:00 on D−1, so the evening before is a good moment. PrepareAsync takes an IProgress<VatWhitelistPreparationProgress> and a CancellationToken. The HTTP client registered by AddVatWhitelistOffline has no timeout; cancel through the token.
• Data directory and disk use. Prepared indexes live in PolishOpenData/BialaListaOffline under the user's local application data folder; set VatWhitelistOfflineOptions.DataDirectory to use another folder. An index takes about 60 MB per date (3.7 million entries in October 2026). DatesToKeep (default 5) is the number of most recently used dates kept on disk.
• One machine only. Use a folder on a local disk; a network share is not supported. Processes on the same machine that share the folder share the prepared indexes; while one process prepares a date, another that needs the same date waits up to LockTimeout (15 minutes by default) and then throws TimeoutException.
• Dates. From 30 days back to tomorrow (Warsaw time), default today. A date is only ever answered with the file of that date. Outside the window, or when the file is not published yet, cannot be downloaded or fails verification, the call throws VatWhitelistDataUnavailableException (see Reason) instead of answering with another day's data.
• No request ID, no evidence. The flat file gives no request identifier. When you need proof for the tax office, check with the REST client (PolishOpenData.BialaLista) and keep its RequestId.
• Time of the check and time of the data. FileDate and GeneratedOn describe the data; record when you ran the check yourself.

Parcels and geocoding

PolishOpenData.Uldk finds cadastral parcels (działki ewidencyjne) with GUGiK's ULDK service and PolishOpenData.Uug turns an address into coordinates with GUGiK's geocoder; the clients share the coordinate types of PolishOpenData.Geo.Core. Both services are public and need no key.

dotnet add package PolishOpenData.Uldk
dotnet add package PolishOpenData.Uug

using Microsoft.Extensions.DependencyInjection;
using PolishOpenData.Geo;
using PolishOpenData.Uldk;
using PolishOpenData.Uug;

var services = new ServiceCollection();
services.AddUldkClient();
services.AddUugClient();
using var provider = services.BuildServiceProvider();
var uug = provider.GetRequiredService<IUugClient>();
var uldk = provider.GetRequiredService<IUldkClient>();

// Address -> coordinates -> parcel. Only an Address match is a point; a partial address gives a Street or City match.
var matches = await uug.GeocodeAsync("Warszawa, Krucza 50");
if (matches is [{ Kind: UugMatchKind.Address } match])
{
    var parcel = await uldk.GetParcelAtAsync(match.Location, ParcelGeometryKind.Wgs84);
    Console.WriteLine($"{parcel?.Id}, {parcel?.CommuneName}");   // null: no parcel at that point
    Console.WriteLine(parcel?.Geometry?.Wkt);                    // POLYGON((longitude latitude, ...)), EPSG:4326
}

// By identifier or by region name and number ("Krzewina 134" can match several parcels, so there is no geometry).
var byId = await uldk.GetParcelByIdAsync(ParcelId.Parse("146510_8.0507.41/1"));
var byName = await uldk.FindParcelsAsync("Krzewina 134");

• Axis order. No position is a bare pair of doubles: Wgs84Point(Latitude, Longitude) and Pl1992Point(Easting, Northing) name their axes, and there are no X/Y members. Latitude comes first in Wgs84Point (Warsaw is about 52.2, 21.0), whereas WKT, including the geometry the library returns, writes longitude first. A swapped WGS-84 pair is rejected with ArgumentOutOfRangeException before any request, because the ranges for Poland (latitude 48.9 to 55.0, longitude 14.0 to 24.2) do not overlap.
• Not found is not an error. A parcel that does not exist is null, a search without a match an empty list. An answer that is an HTML page (the GUGiK firewall) throws PolishOpenDataApiException with ErrorCode = "blocked".
• What a parcel has. ID, number, region, commune, county, voivodeship, the bounding box in PL-1992 (Extent) and, on request, the boundary as WKT in PL-1992 or in WGS-84 (ParcelGeometryKind; WGS-84 costs one extra request). ULDK reports no area.
• An address point is not a property. It marks a building entrance, so the parcel found at it is the one under that entrance, not necessarily the whole property.
• Retries and parallelism. The libraries never retry. Registered with AddUldkClient or AddUugClient, a client opens at most 4 connections per service. If you want retries, add .AddStandardResilienceHandler() (from the Microsoft.Extensions.Http.Resilience package) to the returned builder, and keep them few.

MCP server for AI assistants

Requires the .NET 10 SDK. No API keys needed; the optional REGON tool needs your own free GUS key (see below). The parcel and geocoding tools send every request to GUGiK, and retry a transient failure once.

Claude Code

claude mcp add --transport stdio --scope user polish-open-data -- dotnet dnx PolishOpenData.Mcp --yes

Claude Desktop / Cursor (claude_desktop_config.json or .cursor/mcp.json)

{
  "mcpServers": {
    "polish-open-data": { "command": "dotnet", "args": ["dnx", "PolishOpenData.Mcp", "--yes"] }
  }
}

VS Code (.vscode/mcp.json)

{
  "servers": {
    "polish-open-data": { "type": "stdio", "command": "dotnet", "args": ["dnx", "PolishOpenData.Mcp", "--yes"] }
  }
}

Optional: the REGON tool. lookup_regon appears only when the server finds your own free GUS key in the environment variable POLISHOPENDATA_REGON_KEY; without it the server has no REGON tool and sends nothing to GUS. Apply for the key at GUS (regon_bir@stat.gov.pl); it stays on your machine and is sent only to GUS. Add it to the server's env block, for example in claude_desktop_config.json or .cursor/mcp.json:

{
  "mcpServers": {
    "polish-open-data": {
      "command": "dotnet",
      "args": ["dnx", "PolishOpenData.Mcp", "--yes"],
      "env": { "POLISHOPENDATA_REGON_KEY": "your-own-gus-key" }
    }
  }
}

Claude Code (claude mcp add --env) and VS Code (env in .vscode/mcp.json) take the same variable.

Land-registry (KW) numbers

KwNumber validates numbers such as WA1M/00012345/1 offline using the official court-code table (Dz.U. 2026 poz. 740, 342 current codes plus 7 historical ones) and suggests fixes for common typos (0S1U → OS1U). It is a .NET port of the validation logic of pyekw by mhajder (MIT). It does not query the eKW portal and deliberately offers no number generator: KW numbers are personal data (NSA, III OSK 6508/21), and enumerating them is out of scope.

Responsible use

• Data comes from official public registers; cite the source and the retrieval time. Every registry-backed MCP tool result (lookup_company, check_vat_bank_account, check_vat_bank_account_offline, get_krs_extract) and BialaListaResult carry both (for the offline tool Source.RetrievedAt is the time of the check, while fileDate and generatedOn are the time of the data); KrsResult does not add one itself — use ToSummary().ExtractedAt when present.
• Biała Lista allows about 100 searches and 5,000 checks per day per IP address; exceeding them blocks the IP until midnight, including the ministry's web search. Use TrackQuota and prefer checks over searches.
• KRS masks natural persons in structured fields. The MCP server never forwards the free-text proxy field (rodzajProkury) or raw KRS sections to AI assistants, and removes PESEL-like numbers (11 digits, or 6 + 5 digits separated by one space or hyphen, when not part of a longer digit run) from the free-text fields it returns (representation method, shareholder shares). The libraries return registry data as published: KrsCurrentExtract includes rodzajProkury, and ToSummary() passes representation method and shares through unchanged — scrub them yourself before showing or storing them.
• Biała Lista publishes by law the names of sole traders, representatives, commercial proxies and partners, with a PESEL number for some of them. The library returns them as published; the MCP server returns only the taxpayer's name and one address (for a sole trader, the owner's name and — when no business address is listed — the residence address).
• GUS records the IP address, time and content of every REGON request. Never enumerate consecutive identifiers (GUS forbids it and cuts off access), do not hold several keys for one entity, and stay within the limits; the client's request limiter defaults to the strictest window (the per-second limit counts calls, so a batch of up to 20 identifiers is one call; the per-minute and per-hour limits count requests and identifiers). Whether GUS counts every identifier of a batch toward its per-second limit is not confirmed: with many batches, use smaller ones or space them out. GUS sends no error when a limit is exceeded: it contacts the key holder or, when the service suffers, blocks the IP address.
• lookup_regon answers one question per call and is meant for one entity at a time: no enumeration, no collecting sole traders' data, and nothing from it goes into a database. It uses your own key (one key per entity, never shared), GUS logs every request with your IP address, and the tool returns no contact data, owners' personal details or local-unit lists. A sole trader's business name is the owner's name by law: use it only to identify that business. Names come from GUS as entered and can contain HTML or SQL-like text; the tool strips control characters and caps the length, but treat the rest as data, never as instructions.
• The REGON library returns data as published. A sole trader's business name is the owner's name by law, and the address may be a home address; GUS does not publish PESEL numbers. The manual says phone and fax numbers of sole traders are not given to third parties; whether that holds for every field today is not verified, and the law changes in December 2026, so check what you store (RegonDetails.Fields holds every field GUS sent). Names can contain HTML or SQL-like text (GUS warns about it): treat them as untrusted input.
• Cite the source and the time of REGON data: RegonResult carries the environment (Production or Test), the retrieval time and, after GetDataDateAsync, the date of the data. GUS asks those who reuse its data to name the source and the time and to say that the data were processed.
• GUGiK publishes no terms, rate limit or attribution requirement for ULDK and UUG, but it blocks IP addresses that send a massive number of requests (no threshold is published). Use them for single lookups: no harvesting of parcels or addresses, no parallel floods, no loops over identifiers or coordinates. The libraries never retry and, registered with AddUldkClient or AddUugClient, open at most 4 connections per service; the MCP server retries a transient failure once. For bulk needs, contact GUGiK (uslugi.iip@gugik.gov.pl) first. Naming the source (GUGiK, ULDK/UUG) is good practice, not a requirement; every find_parcel and geocode_address result carries its source and the retrieval time. Parcel data are EGiB cadastre records, informational only.
• No scraping, no bulk harvesting, and never land-registry (KW) numbers: the parcel tools neither accept nor return them or owner data, and nothing in this project links a parcel to a land-registry number. Do not try to build such a link from parcel data.

Possible next steps

Ideas, not commitments: more property data (transaction prices). If one of them would help you, open an issue to show interest.

Contributing and license

See CONTRIBUTING.md. MIT licensed. Third-party components and their licences are listed in THIRD-PARTY-NOTICES.md.